Resemble Updates & Release Notes

Last quarter I wrote that I'd started thinking about how we fight deepfakes differently, and I want to pick that thread back up, because it's the whole reason for what we shipped in Q2.

For years, the job in this space has been detection: something fake gets out and you work to catch it. That work is essential, and it isn't going anywhere but it's not enough on its own.

That's why we've long treated this as a provenance problem too. Provenance just means proof of origin: marking what's real at the moment it's made, so there's something to check against later instead of guessing after the fact. It's the idea behind PerTh, our audio watermarker which we released in 2023.

This quarter, we had a hard reason to push that thinking further. Under the EU AI Act, starting August 2, any AI system placed on the market has to mark its output in a machine-readable format, from the very first file it generates, with no grace period for new systems. The fines for getting it wrong reach 15 million euros or 3% of global revenue.

The question we spent Q2 on was simple: how do you make sure everything that leaves your systems carries its own proof, in a way that survives the real world?

The watermarker we open-sourced in 2023 just got a major upgrade

Watermarking has been part of how we work for years. PerTh has been available since 2023, when we open-sourced it as an audio watermarker. The community ran with it: hundreds of GitHub stars, enterprises running it in production, developers building on top of it.

On June 24 we shipped the rebuild. PerTh Multimodal takes that same idea and extends it well past audio, marking audio, video, image, and text through a single API. It's also far more robust. The original had four known weak spots, everyday edits like pitch shifting and filtering that could wear the mark down, and we rebuilt the training approach to close them. The mark now holds up through compression, re-encoding, and the ordinary processing content goes through once it leaves your hands.

A watermark woven into the content survives the trip through a social platform, even after the file-level record on the outside gets stripped on upload. We hear the fair objection often, that watermarks are easy to strip, so why bother. RAND, the nonprofit research institute and think tank, argued in 2025 that provenance schemes leaning on the whole internet to cooperate won't hold up on their own. The answer is a mark that stays put, one we can still detect with close to 99.9% accuracy when we go looking for it.

The trap most teams don't see coming

Here's the part most teams adding watermarking miss, and it's really a detection problem.

Say you mark all synthetic speech but leave real human speech unmarked. A detector learning from that mix can pick up the wrong signal and decide the watermark itself is what "fake" looks like. One of our researchers, Nicolas Müller, showed exactly this in a June 2026 paper with a German research institute. A detector that falls for it gets worse at spotting content it hasn't seen, lets a fake through the moment someone strips the mark, and can even flag a real person's voice as fake.

You only catch a trap like that if you build both the watermark and the detector. The team that designs the mark is the same team that trains the detector, so we knew to train it on marked real and marked fake audio together, forcing it to learn the real signs of manipulation instead of leaning on the mark. That is the advantage of keeping both under one roof.

One mark isn't enough for the new rule

The August deadline won't be satisfied by a single mark. The EU's guidance, finalized in June, calls for layers working together: the file-level record, an invisible watermark inside the content, and a log to fall back on. Each layer covers for the others when they fail.

PerTh Multimodal is built to be the durable middle layer. It pairs with C2PA, the industry-standard format for the file-level record, so the two reinforce each other, and it handles all four formats from one model instead of a separate tool for each.

If you're mapping your own content before August, the question is narrow. Once your media leaves your systems and hits the open web, is there still something inside it that proves where it came from? If not, that's the gap to close now.

On the detection side our team was very busy

So that's what we rebuilt on the watermark and the other half of the story is detection. Our detection model reached 99% accuracy against NVIDIA's Magpie speech model after another round of training, and on an independent audio deepfake detection benchmark with Podonos, our model ranked first against 8 systems including the likes of Aurigin.ai, Hive, and Reality Defender.

We also keep researching how synthetic voice is built, because staying the clear leader in audio deepfake detection means knowing the newest generation methods before they show up in the wild. That research is open, and it produced four new models this quarter, all free to use: DramaBox, Chatterbox Nano, Chatterbox Flash, and Chatterbox Multilingual.

Finally, outside of approving our accuracy, our team was invited to speak at various events this past quarter. Zohaib joined a panel at AI Insiders in front of CISOs, security founders, and investors, and by his account every panel and every investor thesis kept circling back to detection. He also spoke at Deutsche Telekom's leadership summit in Berlin as one of their T-Challenge winners, where a customer's own leadership spent the session on the problem we build for: scaling AI without scaling the risk that rides along with it.

And Will Krispin, our Head of Partnerships, joined an Okta event on operational risk for federal agencies, where they spoke about how voice is collapsing as an identity signal and a ten-second clip is now enough to beat call-center verification or impersonate a senior official.

What's next

With a busy quarter behind us, we are still full steam ahead. One thing I'm most excited about is what we are shipping next week. Our research team is deep into doing final testing of our new detection model and the early numbers show accuracy in the high nineties against unseen generators.

If you’d like to learn more about anything I wrote about you can book a demo with our team to see detection and watermarking running on your own media or read our latest Deepfake 101 guide for a plain-language, sharable resource of the deepfake threat vectors and how to protect against them.

Original source

Voice quality in TTS used to be the thing that came up in every conversation, lately it doesn't, the bar has moved to latency. The questions are whether the model is light enough to run on the hardware it actually needs to live on without a cloud dependency propping it up, and what happens when you need thousands of concurrent voice agents and your autoregressive model is the bottleneck.

Chatterbox Nano and Chatterbox Flash are both answers to that. Both are open source, shipping today alongside DramaBox and Chatterbox Multilingual V3, which we released recently. One thing that holds across all of the releases: every model ships with watermarking by default. Whatever you build with them, the provenance is already there.

Chatterbox crossed 13 million downloads on Hugging Face somewhere in the middle of building all of this. It's a good reminder of how much this community actually builds with these models.

Chatterbox Nano

Chatterbox Nano is a 110M parameter TTS model built for the edge. It's small enough to run locally and fast enough that the gap between text and audio disappears.

• 10x realtime on GPU, 3x realtime on CPU using 8 threads
• Paralinguistic tags like [laugh], [sigh], [chuckle], and [cough] work natively
• Voice cloning from a 5-second reference clip, no fine-tuning required
• Watermarking baked in by default

It was trained on the same dataset as Chatterbox Turbo, so it carries the same expressiveness. The provenance signal is in the file the moment it is created. And like everything we are shipping this month, watermarking is baked in by default.

Sample:

Welcome to Support Services. For account assistance, press 1. For billing questions, press 2. For technical support, press 3. To speak with a representative, press 0. To repeat these options, press star.

Sample:

Good afternoon, thanks for calling IT support. I'm Alex, your support agent [chuckle]. Let's take a look at what’s going on with your account. Don’t worry, we’ll sort this out together. Go ahead and tell me what you’re experiencing, and I'll walk you through the fix step by step.

Sample:

Thank you for calling CareLine Medical. I'm Howard, your support agent. Please tell me your name and date of birth so I can get your file.

Chatterbox Flash

Chatterbox Flash is Chatterbox rebuilt on a diffusion-LLM architecture for teams running TTS at scale. Most production TTS today hits a ceiling on throughput that hardware alone cannot fix: autoregressive architectures generate tokens sequentially, you can add more machines but you cannot change the loop. Flash changes the loop itself.

• 2x the speed of our autoregressive baseline on vLLM
• Novel prior-subtraction technique that reduces diffusion steps without hurting output quality
• One of the first TTS models built on dLLM architecture that is actually ready to run in production

We think the prior-subtraction technique generalizes to dLLM TTS broadly, and it's a big part of why this is a production model rather than a research one. If throughput is the thing you are working around right now, this is worth a close look.

What and who these are for

Chatterbox Nano is for builders who need TTS to run where a server cannot follow:

• Voice agents at the edge
• On-device deployments
• Real-time applications where every millisecond is load-bearing

Chatterbox Flash is for teams running inference at volume who have hit the ceiling on autoregressive generation and need a way through it.

Both ship as open source under the MIT license.

How to get started

Model cards and quickstart codes for both are on Hugging Face - Chatterbox Nano and Chatterbox Flash.

Original source

Synthetic media and the EU AI Act

Synthetic media is now good enough that the question "is this real?" has become genuinely hard to answer. The EU AI Act's response to that problem is Article 50: all AI-generated media must be marked in a machine-readable format by 2 August 2026. Research published at ACM in 2025 found that only 38% of AI generators currently implement adequate watermarking. How generative model companies and enterprises operate and how regulators want them to operate is quite far apart.

PerTh Multimodal is how we decided to serve the creators and enterprises that need to watermark audio, video, image, and text coverage with a single API - for their own IP protection and for current and future regulation.

When PerTh shipped in 2023, it was an audio-only watermarker. The community picked it up fast with hundreds of GitHub stars, enterprises running it in production pipelines, and a base of open source users building on top of it. But the team had been documenting its limitations honestly: four attack categories consistently degraded watermark recovery in ways that mattered in production and the coverage was always narrower than the compliance problem it was meant to solve. PerTh Multimodal is the ground-up rebuild, and the rest of this post explains exactly how it fixes each limitation.

This post is written for engineers who want to understand what changed and why the design decisions were made the way they were. If you want the short version: PerTh Multimodal achieves near-100% accuracy on all internal audio metrics; expands coverage to video, image, and text; improves performance for custom watermark payloads; and ships with ONNX support for edge and mobile deployment.

What is PerTh?

PerTh, short for "perceptual threshold", started as a name for a specific idea: embed watermark data inside the regions of audio that sit below what the human ear can detect. For PerTh Multimodal, that name describes a family of models with the same design goal across audio, video, image, and text: embed an imperceptible, tamper-resistant signature that survives whatever the content goes through after it leaves your pipeline.

Each PerTh modality uses two models bundled together:

• An encoder that adds an imperceptible signature to the input media
• A decoder that recovers that signature from any downstream copy

What differs across modalities is what “imperceptible” means for each medium, and which domain the encoder and decoder operate in.

PerTh Multimodal modality coverage at a glance:

MODALITY | ENCODING APPROACH | EXPLICIT MARK (CUSTOM MESSAGE) SUPPORT
Audio | Spectral / psychoacoustic | Yes
Image | Pixel-level | Yes
Video | Pixel-level | Yes
Text | Linguistic rewriting | No

Text carries a presence signature only — no custom message payload.

How each modality works

Audio: Psychoacoustics as the encoding channel

Human hearing has uneven sensitivity across frequencies, and a louder sound at one frequency creates a masking effect that hides quieter nearby sounds. PerTh places watermark data inside those masked regions, the ones a listener's ear never registers.

Both the encoder and decoder operate in spectral space, modifying frequency content directly via a residual additive mask on the magnitude spectrogram. The architecture does this in the frequency domain rather than the waveform domain for several reasons:

• Waveform-level models have high-bitrate output which can increase latency
• Restricting the additive mask to low-frequency spectrogram regions lets PerTh guarantee perfect reconstruction of high-frequency content, something a waveform-level model typically cannot guarantee
• Restricting the additive mask to low-frequency spectrogram regions also ensures the watermark is difficult to remove without distorting the underlying speech content

That encoder-decoder pair comes in two variants depending on what you need to embed.

Implicit models encode a single bit, the decoder predicts whether a watermark is present or absent, no payload argument needed.

Explicit models allow a custom bit-string of fixed length to be embedded and later recovered.

Image and video: Better than open source baselines

The image and video models use a pixel-level encoding approach following similar techniques to Google’s SynthID and Meta’s Videoseal, where the encoder makes imperceptible modifications to pixel values to carry the watermark payload. The decoder knows the pattern to look for and can recover the payload from a modified copy.

The main difference is Google extended SynthID to a small number of invited partners whereas PerTh Multimodal is available to any organisation via API, today, without a selection process. Using a mixture of in-house architectures, data, and methods, we have been able to demonstrate high robustness across a wide range of data augmentations. Benchmark data comparing PerTh Multimodal against a Videoseal baseline is available.

Text: Watermarking through linguistic rewriting

Text watermarking works differently from the signal-level approaches used for audio, image, and video. An encoding model rewrites the input text with subtle changes like different word choices or minor structural variations that carry the watermark pattern. The output says the same thing; the pattern of changes is the watermark. When the text is later submitted to the decoder, it looks for its own linguistic signature to determine whether it authored the content.

What the PerTh V1 Audio Model Got Wrong and How PerTh Multimodal Fixes it

The V1 limitations were specific and documented. Four attack categories consistently degraded watermark recovery:

• Waveform pitch shifting
• Waveform low-pass filters
• Additive noise
• Contiguous spectral band masking / Notch filters

None of these are exotic attacks — pitch shifting is standard in audio production, low-pass filtering is applied by default in many encoding pipelines, and additive noise is present in the majority of real-world recordings. An attacker who understood V1's architecture could apply any of them and meaningfully degrade watermark recovery without destroying the underlying audio.

PerTh Multimodal was rebuilt to address each failure mode through a completely new training attack system and a redesigned training curriculum. For image and video, the augmentation suite covers spatial and colour-domain transformations including brightness, colour jitter, crop, blur, JPEG compression, resizing, and many more transformations most likely to be applied to visual content in production environments.

The PerTh Multimodal Training Attack System

The fundamental challenge in building a robust watermarker: the decoder has to be trained on augmented (“attacked”) data, as well as clean data. Before the decoder sees the watermarked signal, augmentations are applied (compression, filtering, pitch shifting, re-encoding, and many others) and the decoder learns to recover the watermark from a signal that has already been degraded.

The constraint that makes this technically hard: those augmentations must be differentiable, or approximately so, for gradients to flow back through training. Arbitrary signal processing cannot be used as a training augmentation without careful design.

PerTh Multimodal solves this with a library of stable augmentations built for the spectral, waveform, and image domains, which all produce usable gradients or have stable dynamics with straight-through estimation.

Examples of spectral attacks used:

• Gaussian Noise
• Scaling
• Low-clip / High-clip
• Quantization
• Band Masking
• Time Masking
• Reverb

Examples of waveform attacks used:

• Precision Reduction
• Mu-law Compression
• Dithering
• Gaussian Noise
• Wavelet Noise

Examples of image/video attacks used:

• Various Codec / Compression Round-trips
• Rotation, Scaling, Stretching, Cropping, etc
• Quantization
• Noising (Salt & Pepper, Gaussian, etc)
• Filters / Effects (Brightness, Shadows, Sharpness, etc)

This is a subset of the full augmentation library, which is constantly expanding.

The breadth matters because diversity of training attacks is how robustness generalizes. Researchers benchmarking watermarkers at Interspeech 2025 found that neural compression posed the most significant challenge even for models trained with compression augmentation and that attack diversity improves robustness, but is insufficient on its own. PerTh Multimodal’s design responds to exactly that finding: attack diversity plus curriculum design, not either alone.

The Training Curriculum: Hard-Example Mining and Adaptive Batching

Attack diversity gets you coverage. What determines whether the model actually generalises to hard attacks is how the training curriculum evolves once those attacks are in the mix.

Standard training applies augmentations with fixed weights throughout the run. Once a model gets accurate on easy attacks, it keeps seeing them at the same rate as hard ones.

PerTh Multimodal introduces train-time hard-example mining. The model is periodically evaluated against held-out data, with accuracy measured per attack type. Attack categories with poor evaluation accuracy get higher weights in the random augmentation selection. The training distribution adapts to wherever the current weaknesses are.

Adaptive batch scheduling runs alongside this. Larger batch sizes late in training produce smoother gradients and improve convergence. PerTh Multimodal automates this with a batch_schedule hyperparameter that increases batch size over time, automatically configuring dataloaders and gradient accumulation within a defined VRAM budget. Learning rate scheduling was removed — smoother gradients from larger batches made it redundant, and it was observed to introduce instability in some training runs.

Decoding Thresholds: What the Numbers Mean for Your Integration

Watermark decoding is probabilistic. Under ideal conditions, a 16-bit PerTh audio watermark has a bit recovery probability above 97%. The practical question for integration is: how many bits need to match before declaring a watermark detected?

That threshold is a tradeoff between detection rate and false positive rate. Let ‘k’ represent the number of bits that must match to declare a watermark detection.

Threshold | Detection Rate | Random False Positive Rate | Best For
k=13 | 99.9% | ~1% | High-recall pipelines: provenance logging, compliance
k=14 | 98.9% | ~0.2% | High-precision pipelines: legal evidence, rights enforcement
k=15 | 91.8% | ~0.03% | Very low false positive tolerance

k = the number of bits that must match to declare a watermark recovered; a higher k lowers the false positive rate.

The table above shows possible threshold choices, and the trade-offs that result from them. Two setups cover most real-world use cases:

Threshold 13 is the recommended default. 99.9% detection rate, roughly 1% chance a random bit string triggers a false positive. The right choice when missing a real watermark is the higher cost.

Threshold 14 is better when false positives carry legal or regulatory consequences such as rights enforcement filings or court-ready evidence packages.

There are two main limitations to understand upfront:

1. Traditional digital signatures are not compatible with probabilistic decoding. PerTh recovers a random subset of bits, which limits cryptographic constructions. The watermark message payloads are effectively public information. PerTh's security model rests on making watermark removal difficult, not on hiding the payload contents. For deployments requiring strong cryptographic ownership guarantees, a custom per-user model is the appropriate solution.
2. Attack robustness is much lower for the explicit 16-bit watermark model. Under some conditions, per-bit accuracy may be below 80% on average. The implicit (1-bit) watermarker does not suffer from this limitation.

Edge Deployment: ONNX Built In

Because PerTh Multimodal’s modules are compatible with standard ONNX op-sets, converting for mobile or edge deployment is straightforward across all modalities. The implicit audio model variant has been validated down to 500ms input chunk sizes, covering streaming and real-time edge deployment. The explicit audio model variant requires larger chunk sizes due to the higher message bitrate, so the choice of chunk size impacts performance and accuracy. For the explicit model, chunk sizes up to 2 seconds are required for high accuracy.

What the EU AI Act Means for Teams Evaluating Watermarking Right Now

Article 50 of the EU AI Act requires that AI-generated audio, video, image, or text be marked in a machine-readable format, with full enforcement from 2 August 2026. The December 2025 draft Code of Practice moves from principle-based guidance to specific technical requirements, mandating a multilayered approach that includes imperceptible watermarks embedded into content and surviving compression and editing.

The Commission's March 2026 guidance frames those requirements in engineering terms: solutions must be "effective, interoperable, robust, and reliable." The training attack system is what makes the robustness requirement concrete — diverse augmentations that push the decoder to generalise rather than memorise. ONNX support handles interoperability. The threshold table above gives compliance teams the quantitative handles to demonstrate reliable performance to regulators.

PerTh Multimodal also pairs with C2PA for dual-layer provenance: C2PA handles file-level metadata (who created it, when, with which system), while PerTh Multimodal handles the signal-level watermark that survives re-encoding and format conversion after metadata has been stripped. A March 2026 paper highlighted potential conflicts between C2PA manifests and invisible watermarks when treated as competing mechanisms, so using them as complementary layers (which is how PerTh is designed) avoids that conflict.

Evaluating PerTh Multimodal for a compliance use case? Explore Resemble Watermarker.

What Has Not Been Evaluated Yet

Three areas remain under active evaluation:

Non-speech audio.

PerTh Multimodal is trained and evaluated on speech datasets. Background music and environmental audio is out-of-distribution for current models. Generalization to non-speech content is untested.

Real-world case studies of explicit watermarking.

Given the limitations in decoding performance, we are looking for parties interested in evaluating the explicit watermarking model. This would provide more clear data surrounding accuracy and use-cases.

Getting Started

View the API documentation at https://docs.resemble.ai/detect/watermark

For teams evaluating PerTh for a production pipeline or EU AI Act compliance use case, explore Resemble Watermarker or talk to the team.

Original source

Why Watermark Voice AI Outputs

Voice AI has reached the point where listener-level detection of synthetic speech is unreliable for most languages. We can no longer definitively say "this sounds synthetic" vs "this sounds real" by ear alone, and our own evaluation work for this release is part of how we know it.

Today, June 10, 2026, we are releasing Chatterbox Multilingual v3 following the launch of an NVIDIA NIM - now available in the NVIDIA build catalog.

Chatterbox Multilingual v3 is the next general-purpose multilingual TTS model in the Chatterbox family. It supports 25 total languages (including 4 dialects and 6 tuned Language Pack models), keeps the same 0.5B Llama backbone as v2 under MIT license, and ships meaningful improvements in speaker similarity, hallucination rate, and conversational naturalness.

Chatterbox Multilingual v3 open source ships with PerTh watermarking embedded by default on every self-hosted audio output, across every supported language. Provenance and the ability to automatically watermark outputs is a major component of this release as regulation like the EU AI Act Article 50 loom, and as voice AI and text-to-speech models strive to be more trustworthy and verifiable at generation time.

This post covers the watermarking architecture first, then walks through the v3 model improvements, the language-by-language evaluation, the NVIDIA NIM deployment path, and the limitations we know about.

Every audio file generated by v3 and the Single-Language Pack is PerTh watermark ready. PerTh is a neural watermarking system that encodes an imperceptible signal into the audio waveform at generation time. It is designed to survive the manipulations that strip metadata-based provenance: MP3 and Opus compression, telephony codecs, audio editing, format conversion, partial clipping, and the resampling that happens when audio moves between platforms. The watermark is detectable with near 100% accuracy on unmodified outputs and degrades gracefully under manipulation.

We made the decision early in Chatterbox's development to ship watermarking by default in the open source repo, and that decision matters more for v3 than it did for earlier releases. v2 was good enough that synthetic outputs were already difficult to distinguish from real recordings without forensic tool results.

The v1 Hugging Face TTS Arena results were the first signal that we were past the listener-level distinguishability threshold on English). v3 is meaningfully better than v2. The ability to confidently state "this sounds synthetic" vs "this sounds real" is now low enough on most supported languages that listener-level detection is unreliable, which puts more pressure on technical provenance signals like watermarking.

On-by-default watermarking, applied at the final stage of text-to-speech generation, is what produces a useful provenance signal at the scale audio AI now operates at. The argument is not that detection will fail, it is that provenance and detection together cover more of the threat surface than either alone, and the marginal cost of embedding a watermark at generation time is small enough that we see no reason to leave it off.

The PerTh detection library is on GitHub under permissive license, so any developer integrating Chatterbox can verify provenance downstream without paying for or licensing additional tooling. To add watermarking to Chatterbox Multilingual outputs deployed via NVIDIA NIM, contact us for a Watermarker API key.

Watermarking + coming regulation

Regulators have started to converge on the answer that the field has been converging on for a few years. The EU AI Act Article 50 requires that AI-generated audio be marked in a machine-readable format starting in August 2026. The TAKE IT DOWN Act signed into law in the US last year creates new content-traceability requirements for synthetic media. PerTh was designed before either of those frameworks existed, but the design choices that made it useful in 2023 (imperceptibility, robustness to manipulation, frame-level encoding) are the same ones that make it compliant with the emerging regulatory environment.

What's new in v3

Chatterbox Multilingual v3 uses the same 0.5B Llama-based backbone as v2. The improvements come primarily from data, training, and language-specific specialization rather than from any increase in model size.

For v3, we expanded and rebalanced the multilingual training mixture with a stronger focus on high-quality, expressive, conversational data. The data recipe places more emphasis on priority languages and regional variants where users need stronger pronunciation control and more stable speaker transfer. We expanded the multilingual training mixture from 25.6k hours to 36.7k hours, and improved balancing across script reading, narration, and conversational data to better match real-world usage patterns we observed in the v2 deployment.

The training updates were aimed at the main reliability issues we observed in earlier multilingual checkpoints. We targeted off-prompt continuation, repetition, accent drift, and degraded speaker similarity in cross-language voice cloning specifically. v3 uses a cleaner and more controlled fine-tuning regime with fine-grained data labeling, stricter sample filtering, and language-specific refinement for the Single-Language Pack. The Single-Language Pack is designed to complement the general multilingual model by giving selected languages and dialects dedicated capacity and tighter quality control.

The result is a model that is more stable, more expressive, and more conversational while keeping the same 0.5B footprint. In our internal testing, v3 shows improved speaker similarity across languages, fewer hallucination-style failures, and more natural delivery for both long-form narration and short voice-agent prompts. v3 is a stronger general-purpose multilingual model for developers building voice agents, creator tools, audiobook workflows, and multilingual speech applications. For priority languages where regional pronunciation or specialized behavior matters most, the Single-Language Pack provides dedicated checkpoints.

Languages supported in v3

1. Arabic (ar)
2. Chinese (zh) - Language Pack
3. Czech (cs)
4. Dutch (nl)
5. English (en)
6. Finnish (fi)
7. French (fr)
8. German (de)
9. Hebrew (he)
10. Hindi (hi) - Language Pack
11. Italian (it)
12. Japanese (ja)
13. Korean (ko)
14. Norwegian (no)
15. Polish (pl)
16. Portuguese (pt)
17. Portuguese Brazilian (pt-br) Language Pack - Dialect
18. Portuguese Portugal (pt-pt) Language Pack - Dialect
19. Russian (ru)
20. Spanish (es)
21. Spanish Mexico / LATAM (es-mx) Language Pack - Dialect
22. Spanish Spain (es-es) Language Pack - Dialect
23. Swedish (sv)
24. Turkish (tr)
25. Vietnamese (vi)

How we evaluated v3

We measured v3 across all supported languages using a transcription-based Character Error Rate (CER) protocol. For each language, we generated 100 samples from held-out prompts, transcribed the synthesized audio with whisper-large-v3 set to the target language, and computed CER against the source text. The intuition is straightforward: if a TTS model hallucinates content, drifts into another language, or repeats itself, an ASR system attempting to transcribe its output back to text will introduce character-level errors that compound visibly. Lower CER on this protocol means the model is saying what it was asked to say.

We chose CER rather than the more common Word Error Rate because it is more sensitive to the failure modes we most wanted to address from v2: small mispronunciations, partial repetitions, and character-level drift in languages where word boundaries are ambiguous, like Chinese and Japanese. WER tends to mask these failures because a single garbled syllable inside an otherwise correct word can register as a full word error or no error depending on tokenization.

CER is still an imperfect proxy for what we actually care about. It captures intelligibility and stability, but it does not measure prosody, speaker similarity, or expressive delivery, all of which matter for the conversational and voice-agent use cases v3 is built for. The results in this post should be read as evidence of v3's intelligibility and stability, not as a complete picture of quality. We are working on an extended evaluation suite that adds subjective Mean Opinion Score (MOS) ratings and objective speaker similarity benchmarks, and we will publish those numbers when they are ready.

Benchmark Results

CER by Language

The strongest pattern in the data is at the top. Italian and German base v3 land under 0.20% CER, and the Single-Language Pack variants for Latin American Spanish (0.28%), Mandarin Chinese (0.41%), and European Spanish (0.55%) sit alongside them in the sub-1% band. English (0.65%) and Hebrew (0.93%) round out the languages where v3 reaches the measurement floor of our transcription-based protocol. For all of these, the next round of improvements will need to come from prosody and speaker-similarity work rather than from reducing transcription errors.

The middle band is roughly 1% to 5% CER and covers Dutch, Russian, base Spanish, European Portuguese, Brazilian Portuguese, base Portuguese, Norwegian, and Turkish. These are languages where v3 is production-ready for general use, with the dialect-specific Single-Language Pack models pulling Spanish from 2.89% base to 0.28-0.55% in the dialect variants. Portuguese shows a more modest improvement from the base model (3.85%) to the dialect packs (3.01-3.77%), which is consistent with the underlying point about specialization: the headroom for the dedicated models is largest where the base model has the most distance left to travel.

Hindi, French, Polish, and Arabic land in the 6-8% range, which is meaningfully worse than the front-runners but still usable in production with appropriate testing. Hindi is in the Single-Language Pack at 2.55 - 6.72% CER, which is roughly where the base Hindi performance was in earlier checkpoints. The dedicated capacity is doing the work of preventing degradation rather than driving large gains, and we are continuing to invest in Hindi-specific data work for future releases.

The bottom of the table tells a different story, and we discuss the languages that did not reach production quality in the limitations section below.

Results — Latency and deployment

On a single H100 with our unoptimized PyTorch implementation, v3 measures the following on our internal benchmark:

For voice-agent and conversational deployments, TTFB is the number that matters: it determines how quickly the first audible response reaches the user. Under 300ms is at the threshold of what feels like natural conversational pacing, which v3 clears with headroom. The RTF of approximately 5 means v3 generates audio roughly five times faster than real time on a single H100, leaving room for batching multiple concurrent streams on the same GPU.

These are the unoptimized PyTorch numbers. The NVIDIA NIM release covered in the next section ships with TensorRT-LLM acceleration, optimized batching, and dynamic shape support, which lower TTFB and raise concurrent throughput substantially relative to this baseline. This means quicker round trip latencies when deployed as part of a voice AI agent, or faster responses in a one way use case where the voice output is the end goal.

For Enterprise-Scale Deployment: NVIDIA NIM Available

NVIDIA NIM, is an optimized, containerized inference endpoint with TensorRT-LLM acceleration, of the model. In the table below, you can see NVIDIA NIM delivers 2x-39x throughput vs unoptimized PyTorch while supporting enterprise-scale concurrent inference.

NVIDIA NIM provides optimized batching, autoscaling, and the deployment plumbing that enterprise voice-agent infrastructure typically needs.

For Chatterbox v3 specifically, that means meaningfully lower latency than the unoptimized PyTorch numbers in the table above, dynamic batching for concurrent streams, and a deployment path that fits cleanly into existing NVIDIA enterprise stacks.

The way we think about the choice between the two: use the open-weight model when you want full control over the deployment, are running at moderate volume, or are doing research, fine-tuning, or evaluation work. Use NVIDIA NIM for production when you want:

• Full control over the deployment
• Security patches on the NIM container
• Free API for evaluation available with build.nvidia here

All at the scale where the optimization and operational plumbing pay for themselves. Both options ship from the same v3 weights and produce equivalent output quality. The difference is in the inference layer and how much of the production engineering you want to own.

Why ship a general model and a Single-Language Pack at the same time

The architectural decision behind this release was not technical, it was about how developers actually deploy multilingual TTS. We started with the assumption that one good general model should be enough, but the v2 deployment data disputed that assumption in two specific ways.

First, the highest-volume single-language users of v2 were not deploying it in English. They were deploying it in Hindi, Brazilian Portuguese, and Latin American Spanish, and they were running into quality ceilings that the general multilingual model could not break through no matter how much fine-tuning they did at the application layer. These were languages where v2 had measurable headroom against what a dedicated model could achieve, and where the deployment economics justified a larger, language-specific model.

Second, the dialect question turned out to matter more than we initially expected. European Portuguese and Brazilian Portuguese share most of their grammar and a large fraction of their vocabulary, but they have substantially different phonologies, and Portuguese-speaking listeners can distinguish them effortlessly. A model trained on broadly-balanced Portuguese data produces speech that sounds technically correct but slightly off to native ears of either variant. The same applies, less dramatically, to European Spanish and Latin American Spanish. The Single-Language Pack lets us serve each variant from a model that has only ever seen that variant, which is the closest we can get to native-sounding output without language-specific architectural changes.

The tradeoff is parameter efficiency. Six dedicated models cost more to host than one general model, and they only make sense when the deployment volume justifies the additional infrastructure. We expect most developers will use v3 base for breadth and reach for a Single-Language Pack model only when the language matters enough commercially to be worth a dedicated GPU footprint.

Model Cards on Hugging Face

CER % = Character Error Rate; lower is better.

What v3 still does not do well

v3 has known weaknesses we want to surface in detail rather than leave them to be discovered by developers in production.

Korean and Vietnamese are not at production quality.

v3 generates Korean and Vietnamese speech with CERs of 70.90% and 75.21% respectively, both more than an order of magnitude worse than the next-weakest language in the supported set. The root cause is training data: we do not have sufficient volume of high-quality paired text-audio data in either language across the speaker distribution the model needs to generalize properly. v3 will produce Korean and Vietnamese output, but we recommend against using v3 for commercial deployments in either language until we have shipped an improvement. We are actively working on the data side and expect to release either upgraded multilingual checkpoints or dedicated Single-Language Pack models as follow-ups.

Czech, Japanese, and Finnish are usable with caveats.

Czech at 32.21% CER, Japanese at 21.87%, and Finnish at 17.52% are all behind the front-runners but still in the range where the output is intelligible for many applications. We attribute the gap to a combination of factors: smaller training data volume than the well-supported European languages, tokenization challenges (particularly for Japanese mixed scripts of hiragana, katakana, and kanji), and prosodic features our current evaluation does not capture well. Developers building applications in any of these three languages should evaluate carefully on their specific content before deploying.

Watermarking has limits we are continuing to characterize.

PerTh is robust against the manipulations we have tested, but we have not yet published systematic robustness curves under adversarial conditions (specifically: targeted attacks designed to strip the watermark without destroying audio quality). The watermark also produces a binary signal for "this audio was generated by Chatterbox," not a model-version signal or a tenant-attribution signal. These capabilities and on their way, and we are working with the C2PA community on the interop standards that would let multiple TTS providers' watermarks be verified through a single library.

The deeper limitation is that CER does not capture everything.

A model can score well on transcription-based metrics while still missing prosody, sounding unnatural, or failing on speaker similarity in voice-cloning use cases. The numbers in this post are evidence of intelligibility and stability, not a complete picture of quality. We are working on the next round of evaluations and expect those numbers to surface tradeoffs the current results hide.

What's next

The work continues in four directions. The first is closing the gap on the languages where v3 is weak, particularly Korean and Vietnamese, which will require both data work and likely architectural attention for languages with complex tokenization. The second is extending the evaluation protocol to add subjective quality and speaker similarity metrics that the current CER-based approach misses. The third is broadening the Single-Language Pack to additional languages where the deployment volume and quality headroom justify dedicated models. The fourth is the watermarking work: tenant attribution, model-version signaling, and adversarial robustness characterization.

Original source

Directable speech

At some point, everyone who has built with TTS hits the same wall. The voice sounds fine, the words are correct, and it still sounds wrong in a way that is hard to explain until you have heard it enough times to name it: flat. Technically accurate, completely unconvincing.

The problem was that we were using machine language to describe something fundamentally human. Tags, parameters, style tokens, all of it was a translation layer between what you actually wanted and what the model could understand. The translation is where the performance struggled to sound…human.

A real director does not say "increase emotional intensity by 20%." They set a scene, they describe a character, and the performance follows from that direction. DramaBox is the first TTS model that works the same way. You describe what you want in plain language, the way you would describe it to a person, and the model understands it that way.

The second problem was quieter but just as serious. Every piece of synthetic audio left your hands with no way to prove it was yours. No signal in the file, no chain of custody, nothing that holds up when legal gets involved.

Today we are releasing DramaBox, and it addresses both.

Huge congratulations to Resemble AI on the release of DramaBox, their latest expressive voice model. It's also amazing to see the Resemble TTS family surpass 10M downloads on Hugging Face, a testament to the strength of the open model community.

• Joshua Lochner, Open Source Machine Learning Engineer

The format works like a screenplay. Speaker description and stage directions go outside the quotes. Dialogue goes inside. The model speaks the dialogue and interprets everything outside the quotes as performance direction, never as words to say.

Read this prompt:

A regal woman speaks with cold fury in a measured, low voice. She sighs deeply, "I have told you a thousand times, and yet here we are again." Her voice sharpens with rising anger, "Do you honestly think I enjoy repeating myself?!" She drops to a venomous whisper, leaning close, "Now get out of my sight before I do something we will both regret."

That is the exact prompt we passed to the model, one generation with no post-processing and no stitching of takes together.

🎧 Sample:
Regal Queen — Cold Fury to Venomous Whisper

How it works

The prompt format distinguishes between two zones. Inside double quotes, the model speaks literally. Outside double quotes, it performs. Write, She sighs deeply outside the quotes and the model produces a sigh. Write "Sigh" inside the quotes and the model says the word out loud, which is not what you want. That distinction is what makes the whole thing work, and it is also what makes the prompting feel genuinely different from anything else in TTS.

Here is what that looks like when you give the model a character who cannot stop laughing:

A playful girl speaks in a bright, singsong voice, already mid-giggle, "Hehehe, oh my gosh you should see your face right now, it is priceless!" She gasps for air between giggles, "Oh my, I cannot stop laughing!" She tries to compose herself with a long sigh, "Ahhhhh okay okay, I will stop, I promise." She leans in and whispers, "But seriously though, between you and me," then immediately loses it again, "Haha, I just cannot! You are way too funny!"

🎧 Sample:
Catgirl — Uncontrollable Giggling

The gasps, the failed attempts to compose herself, the snort at the end — all of it came from the prompt, not from post-processing.

What's under the hood

DramaBox is a 3.3B-parameter audio-only diffusion transformer, LoRA-merged, with a Gemma 3 12B text encoder running at 4-bit quantization. It outputs 48kHz stereo audio in AAC or WAV. On a warm H100, generation takes around 2.5 seconds.

Voice cloning is optional. Pass a 10-second reference clip and the model clones the target timbre while still following the prompt for everything else: emotion, delivery, the laugh in the middle of a sentence, the breath before the hard line. The reference sets the voice and the prompt directs the performance. Without a reference, the model invents a voice that fits the scene description.

And it’s a step towards compliant TTS for coming regulations like the EU AI Act. Every output is watermarked with Resemble Watermarker. The watermark is embedded in the signal itself, not in file metadata, and constrained to speech-relevant frequencies so it is inaudible to listeners. It survives MP3 and AAC compression, re-encoding, and common edits at ~100% detection accuracy. Pass that output through our deepfake detection API and you get back a binary decode, not a probability score you have to interpret, but a watermark-present-or-not answer that is more defensible when legal gets involved. Generation and protection are not separate decisions here. They are the same decision, made at the moment of creation.

The range

The model handles tonal range across the full dramatic spectrum. A villain whose menace never tips into parody:

🎧 Sample:
Villain — Sinister Laugh

A talk show host who loses it completely and cannot find his way back:

🎧 Sample:
Talk Show Host — Wheezing Laughter

Five voices building a late-90s pop harmony from soft synchronized layers to a full-group chorus:

Backstreet Boys, a polished late-90s boy band with five smooth, harmonizing male voices. "Step by step… out the door… new day… ready for more…" they sing in soft, synchronized harmony. One voice steps forward. "Keys in my hand… got my plan…" The others swell behind him. Their voices rise together. "Tell me why… every morning feels the same…" then "I'm ready to go…" The full group returns in a bright, unified chorus. "We'll make it our way… through the rush, through the noise, we keep moving strong, yeah!"

🎧 Sample:
Backstreet Boys — Pop Harmony

And a football commentator calling a fridge opening with the full gravitas of a Champions League final:

🎧 Sample:
Football Commentator — Martin Tyler

That last one is worth listening to carefully because it demonstrates something more specific than expressiveness: precise tonal control across a long, slow build with crowd audio layered underneath, which is a harder thing to get right than a single emotional peak. The model is not just generating speech. It is generating a scene with pacing.

What and who this is for

DramaBox is an English-only release, and that is intentional. Getting directable speech right in one language is harder than getting flat TTS right in thirty, and we were not willing to trade quality for coverage on this one.

The use cases are anywhere flat TTS has always been the bottleneck: game dialogue that players do not skip, audiobook narration with real character differentiation, voice agents that do not sound like they are reading from a script, dubbing work where the performance has to match the scene and not just the words. Every file that comes out carries a provenance signal embedded at the moment of creation, so wherever it ends up, you can prove it was yours.

Run it self-hosted. The model card and quick-start code are on Hugging Face. If you’re interested in running DramaBox at scale, we suggest you reach out to Cerebrium or GMI.

What's next

DramaBox is the first of four open-source TTS models we are shipping this month, and each one is a different answer to a different problem.

• Chatterbox Nano: 110M parameters. 10x realtime on GPU, 3x realtime on CPU. Runs at the edge. Full paralinguistic tags, voice cloning from 5 seconds. The smallest serious TTS model we've ever shipped, and arguably the fastest open-source TTS.
• Chatterbox Flash: Chatterbox, rebuilt on a diffusion-LLM architecture. 2x faster than our AR baseline on vLLM. Ships with a novel prior-subtraction technique we believe generalizes to any dLLM TTS, one of the first production TTS models on this architecture.
• Chatterbox Multilingual V3: Better speaker similarity, fewer hallucinations, more natural delivery across languages. Plus dedicated single-language models for Mandarin, LATAM Spanish, Brazilian Portuguese, Spain Spanish, Portugal Portuguese, and Hindi.

We will cover each one when it's available.

Original source

When Zohaib and Will came back from MWC Barcelona, the story they told stuck with me.

They ran a game at the booth. Played audio clips to people and asked: real voice or AI-generated? Engineers. Enterprise buyers. Security professionals. People who work in this space every day. Around 70% could not tell the difference.

I keep thinking about that.

I’ve been building in voice AI long enough to know that stat should not surprise me. But it did. Because those are exactly the people who are supposed to catch this.

The gap between real and synthetic has not just narrowed. For most practical purposes, it has closed. And if the people who are supposed to detect it can’t, the rest of us definitely can’t.

So the question we kept coming back to after Barcelona: what do you actually need to operate in a world where you cannot trust what you hear, see, or read anymore?

This is what we built in Q1.

You can't always tell what's real anymore.

I’m thinking about this more as a provenance problem than a detection problem. Not catching fakes after they spread. Proving what you made at the moment you made it.

For a while our watermarking API worked on audio only. That made sense when we first built it. It does not make sense anymore. AI-generated images and video are spreading at the same rate as audio. Conflict zones, boardrooms, political campaigns. We had not kept up.

So in Q1 the team extended it to cover images and video too. The system auto-detects file type, encodes an invisible watermark, and supports sync mode if you need the result right away. One API, three media types, no separate pipelines. Watermarking is how you prove what you made before someone makes something in your name.

[Read the docs →]

Five detection updates, shipped this quarter

Watermarking protects what you generate. But a lot of synthetic media already exists with no provenance at all, and more is being created right now. That’s where detection comes in.

The research and engineering teams shipped five updates to the detection stack this quarter; these are a few that are top of mind.

Accuracy and provenance, together. Image DFD accuracy improved from F1 0.961 to 0.970. But honestly the more interesting update is reverse image search, now live in both the UI and API. Pass use_reverse_search: true in your detection request and you get back provenance data, matches against known debunked hoaxes, and a spread indicator showing how widely the image has already circulated.

This catches things pure ML models miss. Images too new to be in any training set. Synthetic media that has been spreading for days before anyone flags it. If a fake has no digital footprint, the model has nothing to compare it to. Reverse image search gives you a second layer that does not depend on prior training.

Non-face content. We extended coverage to AI-generated content without faces. We shipped this immediately when we saw the type of synthetic conflict imagery circulating on social platforms. The team turned it around in days. This was a case where being behind was not an option.

Three more worth knowing about:

• Identity now surfaces directly in the main Detect app. One request returns both a verdict and the closest matching speaker identity.
• Intelligence Reporting now runs three expert analysis layers in a single structured output: our own model, OpenAI, and Anthropic.
• Voice Activity Detection has been updated to filter non-speech content more cleanly, which means fewer false positives on audio with background noise or music mixed in.

We also shipped Zero Retention Mode. The ask from enterprise and government customers was consistent and non-negotiable: submitted media cannot be viewable after analysis. Not by our staff, not by our systems, not by anyone. Files are purged automatically after processing. If that has been the thing blocking a procurement conversation, it is not the thing anymore.

[Learn More in the Docs →]

On the generation side: custom vocabulary and NVIDIA ACE

While all of this was happening, the team also shipped on the generation side.

Custom vocabulary does not get the attention it deserves. It is not a flashy feature. It is the thing that makes TTS actually usable in healthcare, pharma, and legal. The industries where getting a word wrong is not an inconvenience. It is a compliance problem. I have been pushing for this one for a while.

The old fix was retraining. Slow and expensive. The new fix: pass a short audio sample at inference time and the model corrects the pronunciation on the fly. No retraining, no pipeline changes, correct on the first pass.

[Hear it now →]

And on the quality side: NVIDIA named Chatterbox in their GDC 2026 RTX developer blog as part of NVIDIA ACE, their suite for expressive on-device NPC voice in games. NPC dialogue is historically one of the first things players skip. The bar for voice that actually holds attention is high. Being named in that context matters to us.

Start here

We built detection for enterprises. But I don’t think detection should only be accessible to enterprises.

Try out @resemble_detect bot on X. It’s free, you don’t need to create an account, and there is no API key. Tag it with any image or video and get a result back. It’s the quickest way to try our deepfake detection.

Resemble AI Deepfake Detector (beta). Puts a Scan button on every image, video, and audio element on any webpage. Powered by the same detection technology used by enterprises and government agencies, now free to try. Install it, sign in to your Resemble AI account, and a Scan button appears automatically on media across Twitter/X, Reddit, Instagram, TikTok, LinkedIn, and more.

Each scan returns:

• A verdict: Authentic, AI Generated, or Uncertain
• A confidence score and rationale
• Heatmap visualization, frame-by-frame video breakdown, and per-segment audio scores

This beta release includes 4 free scans per day.

[Get the extension today →]

What's Next?

I’m genuinely proud of what this team shipped in one quarter. The research team turned around several model updates, one of them in direct response to synthetic conflict imagery spreading in real time. Engineers extended watermarking to three media types without breaking the existing API. The team behind the Deepfake Detector and the X bot made detection accessible to anyone with a browser.

None of this is abstract to me. I track it in Linear every day. We still have alot of work to do.

If you want to understand the full scale of what we are building against, the Deepfake Threat Report 2025 is a great place to start.

[Download the report today →]

Original source

Obaid Ahmed

When Zohaib and Will came back from MWC Barcelona, the story they told stuck with me.

They ran a game at the booth. Played audio clips to people and asked: real voice or AI-generated? Engineers. Enterprise buyers. Security professionals. People who work in this space every day. Around 70% could not tell the difference.

I keep thinking about that.

I’ve been building in voice AI long enough to know that stat should not surprise me. But it did. Because those are exactly the people who are supposed to catch this.

The gap between real and synthetic has not just narrowed. For most practical purposes, it has closed. And if the people who are supposed to detect it can’t, the rest of us definitely can’t.

So the question we kept coming back to after Barcelona: what do you actually need to operate in a world where you cannot trust what you hear, see, or read anymore?

This is what we built in Q1.

You can’t always tell what’s real anymore.

I’m thinking about this more as a provenance problem than a detection problem. Not catching fakes after they spread. Proving what you made at the moment you made it.

For a while our watermarking API worked on audio only. That made sense when we first built it. It does not make sense anymore. AI-generated images and video are spreading at the same rate as audio. Conflict zones, boardrooms, political campaigns. We had not kept up.

So in Q1 the team extended it to cover images and video too. The system auto-detects file type, encodes an invisible watermark, and supports sync mode if you need the result right away. One API, three media types, no separate pipelines. Watermarking is how you prove what you made before someone makes something in your name.

[Read the docs →]

Five detection updates, shipped this quarter

Watermarking protects what you generate. But a lot of synthetic media already exists with no provenance at all, and more is being created right now. That’s where detection comes in.

The research and engineering teams shipped five updates to the detection stack this quarter; these are a few that are top of mind.

Accuracy and provenance, together.

Image DFD accuracy improved from F1 0.961 to 0.970. But honestly the more interesting update is reverse image search, now live in both the UI and API. Pass use_reverse_search: true in your detection request and you get back provenance data, matches against known debunked hoaxes, and a spread indicator showing how widely the image has already circulated.

This catches things pure ML models miss. Images too new to be in any training set. Synthetic media that has been spreading for days before anyone flags it. If a fake has no digital footprint, the model has nothing to compare it to. Reverse image search gives you a second layer that does not depend on prior training.

[Reverse Image Search]

Non-face content.

We extended coverage to AI-generated content without faces. We shipped this immediately when we saw the type of synthetic conflict imagery circulating on social platforms. The team turned it around in days. This was a case where being behind was not an option.

Three more worth knowing about:

• Identity now surfaces directly in the main Detect app. One request returns both a verdict and the closest matching speaker identity.
• Intelligence Reporting now runs three expert analysis layers in a single structured output: our own model, OpenAI, and Anthropic.
• Voice Activity Detection has been updated to filter non-speech content more cleanly, which means fewer false positives on audio with background noise or music mixed in.

We also shipped Zero Retention Mode. The ask from enterprise and government customers was consistent and non-negotiable: submitted media cannot be viewable after analysis. Not by our staff, not by our systems, not by anyone. Files are purged automatically after processing. If that has been the thing blocking a procurement conversation, it is not the thing anymore.

[Learn More in the Docs →]

On the generation side: custom vocabulary and NVIDIA ACE

While all of this was happening, the team also shipped on the generation side.

Custom vocabulary does not get the attention it deserves. It is not a flashy feature. It is the thing that makes TTS actually usable in healthcare, pharma, and legal. The industries where getting a word wrong is not an inconvenience. It is a compliance problem. I have been pushing for this one for a while.

The old fix was retraining. Slow and expensive. The new fix: pass a short audio sample at inference time and the model corrects the pronunciation on the fly. No retraining, no pipeline changes, correct on the first pass.

[Custom Vocabulary]

[Hear it now →]

And on the quality side: NVIDIA named Chatterbox in their GDC 2026 RTX developer blog as part of NVIDIA ACE, their suite for expressive on-device NPC voice in games. NPC dialogue is historically one of the first things players skip. The bar for voice that actually holds attention is high. Being named in that context matters to us.

Start here

We built detection for enterprises. But I don’t think detection should only be accessible to enterprises.

Try out @resemble_detect bot on X.

It’s free, you don’t need to create an account, and there is no API key. Tag it with any image or video and get a result back. It’s the quickest way to try our deepfake detection.

Resemble AI Deepfake Detector (beta). Puts a Scan button on every image, video, and audio element on any webpage. Powered by the same detection technology used by enterprises and government agencies, now free to try. Install it, sign in to your Resemble AI account, and a Scan button appears automatically on media across Twitter/X, Reddit, Instagram, TikTok, LinkedIn, and more.

Each scan returns:

• A verdict: Authentic, AI Generated, or Uncertain
• A confidence score and rationale
• Heatmap visualization, frame-by-frame video breakdown, and per-segment audio scores

This beta release includes 4 free scans per day.

[Get the extension today →]

What’s Next?

I’m genuinely proud of what this team shipped in one quarter. The research team turned around several model updates, one of them in direct response to synthetic conflict imagery spreading in real time. Engineers extended watermarking to three media types without breaking the existing API. The team behind the Deepfake Detector and the X bot made detection accessible to anyone with a browser.

None of this is abstract to me. I track it in Linear every day. We still have alot of work to do.

If you want to understand the full scale of what we are building against, the Deepfake Threat Report 2025 is a great place to start.

[Download the report today →]

Original source

Today we're releasing @resemble_detect, a free bot that lets anyone on X check whether an image or video has been AI-generated or manipulated.

We're doing this because social media platforms are flooded with synthetic media and most people have no way to verify what they're looking at. That's a problem we can help solve. Here's how it works, and what this tool is for, and what it isn't.

How It Works

Using @resemble_detect is simple:

1. Find an image or video on X that you want to verify
2. Reply to the post and tag @resemble_detect with the phrase "is this fake?"
3. We'll analyze the content and reply with a visualization (for images) and a confidence score indicating the likelihood that the content is AI-generated or manipulated

That's it. Free, public, no account required beyond your existing X profile.

The detection is powered by Resemble AI's DETECT technology, the same models that rank at the top of independent benchmarks and process millions of verifications for enterprises, governments, and media organizations.

The Problem Is Bigger Than Deepfakes

The conversation around AI-generated content tends to flatten into a simple binary: real or fake, but the reality is messier.

Content exists on a spectrum of authenticity. An image might be entirely AI-generated, or it might be a real photograph with minor edits, or something in between. A real person's face swapped onto another body. Authentic footage with manipulated audio. A genuine video clipped and recontextualized to misrepresent what happened. Text, images, audio, video, all of it can be altered in ways ranging from trivial touch-ups to complete fabrication.

Misinformation works the same way. Sometimes it's entirely invented. Sometimes it's a real story with one detail changed, a date, a name, a number, and that transforms truth into falsehood. The manipulation can be subtle or total, like the recent X posting by the White House.

Our detection technology handles this spectrum. We can identify fully synthetic content, partially manipulated media, and everything in between. We return a confidence score because the answer often isn't binary, and unfortunately, this kind of content is picked up at the speed of a repost.

But there's one dimension where there is no spectrum: consent.

Consent Is Binary

You either have permission to share something or you don't. There's no gradient here, no "partially consensual," no gray area.

This matters because the worst uses of synthetic media aren't about political misinformation or celebrity face-swaps. They're about non-consensual intimate imagery. They're about harassment, exploitation, and abuse. They're about using AI to create content that real people never agreed to, then posting it publicly to humiliate, threaten, or harm them.

We do not and can not verify consent.

What we can tell you is whether content appears to be AI-generated or manipulated. What we cannot tell you is whether the subject of that content agreed to its creation or distribution. That determination requires context, investigation, and often legal process that no API can replicate.

This distinction is critical, we refuse to blur it.

What We Will Not Do

Let us be unambiguous:

We do not condone the creation or distribution of non-consensual synthetic media. Full stop. Whether the content is AI-generated or authentic, if it's shared without consent, it's a violation.

We do not process explicit content through this public bot. The @resemble_detect bot is not a tool for analyzing explicit imagery. If you're trying to use it for that purpose, don't. Our Terms of Service prohibit it. If this is a recurring issue for you personally, please reach out to us at [email protected] and we will determine whether we can process your requested media.

We find the distribution of CSAM repugnant and will cooperate fully with law enforcement. This should go without saying. We're saying it anyway.

We built detection technology because we believe people deserve to know when they're being deceived. We did not build it to enable new forms of abuse. If you're planning to use our tools to harm someone, find another platform, you're not welcome here.

Why We're Releasing This Anyway

We're aware that any tool capable of detecting synthetic media could theoretically be misused. Someone could use detection to "verify" non-consensual content before sharing it. Someone could use a "real" verdict to add false credibility to manipulated media or claim that if it's real, it can be posted.

But we do believe the alternative is worse. Synthetic media is already everywhere on X. Deepfakes are already being posted, shared, and believed. The asymmetry between creation and detection has left most people defenseless, unable to question what they see, unable to verify what they're told.

Giving people a free, accessible way to check content doesn't solve every problem. But it shifts the balance, it creates friction where there was none. It makes "is this real?" a question people can actually answer, instead of something they have to guess at. And just to reiterate, even if something is real or appears authentic, that doesn't equate to consent.

We'd rather build tools that help people navigate this landscape than pretend the landscape doesn't exist.

We All Deserve to Know What's Real

We're releasing this tool because the current state of synthetic media on social platforms is untenable. People are being deceived at scale, and trust in the media and each other is eroding.

This is our attempt to put detection capabilities directly in the hands of the people who need it most.

It won't solve everything. It won't stop bad actors. It won't verify consent or intent or context. But it will let you ask "is this fake?" and get an actual answer.

Try it now: Tag @resemble_detect on any image or video post with "is this fake?"

Read our Terms of Service: https://www.resemble.ai/resemble-ai-x-bot-terms-of-service/

Questions? Contact us at [email protected].

Original source

Today we’re releasing @resemble_detect, a free bot that lets anyone on X check whether an image or video has been AI-generated or manipulated.
We’re doing this because social media platforms are flooded with synthetic media and most people have no way to verify what they’re looking at. That’s a problem we can help solve. Here’s how it works, and what this tool is for, and what it isn’t.

How It Works

Using @resemble_detect is simple:

• Find an image or video on X that you want to verify
• Reply to the post and tag @resemble_detect with the phrase “is this fake?”
• We’ll analyze the content and reply with a visualization (for images) and a confidence score indicating the likelihood that the content is AI-generated or manipulated

That’s it. Free, public, no account required beyond your existing X profile.
The detection is powered by Resemble AI’s DETECT technology, the same models that rank at the top of independent benchmarks and process millions of verifications for enterprises, governments, and media organizations.

The Problem Is Bigger Than Deepfakes

The conversation around AI-generated content tends to flatten into a simple binary: real or fake, but the reality is messier.
Content exists on a spectrum of authenticity. An image might be entirely AI-generated, or it might be a real photograph with minor edits, or something in between. A real person’s face swapped onto another body. Authentic footage with manipulated audio. A genuine video clipped and recontextualized to misrepresent what happened. Text, images, audio, video, all of it can be altered in ways ranging from trivial touch-ups to complete fabrication.

Misinformation works the same way. Sometimes it’s entirely invented. Sometimes it’s a real story with one detail changed, a date, a name, a number, and that transforms truth into falsehood. The manipulation can be subtle or total, like the recent X posting by the White House.
Our detection technology handles this spectrum. We can identify fully synthetic content, partially manipulated media, and everything in between. We return a confidence score because the answer often isn’t binary, and unfortunately, this kind of content is picked up at the speed of a repost.
But there’s one dimension where there is no spectrum: consent.

Consent Is Binary

You either have permission to share something or you don’t. There’s no gradient here, no “partially consensual,” no gray area.
This matters because the worst uses of synthetic media aren’t about political misinformation or celebrity face-swaps. They’re about non-consensual intimate imagery. They’re about harassment, exploitation, and abuse. They’re about using AI to create content that real people never agreed to, then posting it publicly to humiliate, threaten, or harm them.
We do not and can not verify consent.
What we can tell you is whether content appears to be AI-generated or manipulated. What we cannot tell you is whether the subject of that content agreed to its creation or distribution. That determination requires context, investigation, and often legal process that no API can replicate.
This distinction is critical, we refuse to blur it.

What We Will Not Do

Let us be unambiguous:
We do not condone the creation or distribution of non-consensual synthetic media. Full stop. Whether the content is AI-generated or authentic, if it’s shared without consent, it’s a violation.
We do not process explicit content through this public bot. The @resemble_detect bot is not a tool for analyzing explicit imagery. If you’re trying to use it for that purpose, don’t. Our Terms of Service prohibit it. If this is a recurring issue for you personally, please reach out to us at [email protected] and we will determine whether we can process your requested media.
We find the distribution of CSAM repugnant and will cooperate fully with law enforcement. This should go without saying. We’re saying it anyway.
We built detection technology because we believe people deserve to know when they’re being deceived. We did not build it to enable new forms of abuse. If you’re planning to use our tools to harm someone, find another platform, you’re not welcome here.

Why We’re Releasing This Anyway

We’re aware that any tool capable of detecting synthetic media could theoretically be misused. Someone could use detection to “verify” non-consensual content before sharing it. Someone could use a “real” verdict to add false credibility to manipulated media or claim that if it’s real, it can be posted.
But we do believe the alternative is worse. Synthetic media is already everywhere on X. Deepfakes are already being posted, shared, and believed. The asymmetry between creation and detection has left most people defenseless, unable to question what they see, unable to verify what they’re told.
Giving people a free, accessible way to check content doesn’t solve every problem. But it shifts the balance, it creates friction where there was none. It makes “is this real?” a question people can actually answer, instead of something they have to guess at. And just to reiterate, even if something is real or appears authentic, that doesn’t equate to consent.
We’d rather build tools that help people navigate this landscape than pretend the landscape doesn’t exist.

We All Deserve to Know What’s Real

We’re releasing this tool because the current state of synthetic media on social platforms is untenable. People are being deceived at scale, and trust in the media and each other is eroding.
This is our attempt to put detection capabilities directly in the hands of the people who need it most.
It won’t solve everything. It won’t stop bad actors. It won’t verify consent or intent or context. But it will let you ask “is this fake?” and get an actual answer.

Try it now: Tag @resemble_detect on any image or video post with “is this fake?”
Read our Terms of Service: https://www.resemble.ai/resemble-ai-x-bot-terms-of-service/
Questions? Contact us at [email protected].

Original source

Resemble AI Generative AI Voice-Based Deepfake Simulation Platform for Public Sector

AI-powered threats are no longer science fiction. From voice phishing scams to real-time social engineering attacks, deepfakes have entered the mainstream — and government agencies are among the most at risk.
That’s why we’re excited to announce a major milestone: Resemble AI has launched its Generative AI Voice-Based Deepfake Simulation Platform, now available to the Public Sector through Carahsoft Technology Corp., The Trusted Government IT Solutions Provider®.
This collaboration makes it possible for agencies across federal, state, and local levels to proactively defend against AI-driven threats.
Resemble AI’s platform equips agencies with critical tools and training, decreasing the risk of deepfake attacks. It provides agencies with a crucial protection layer and the ability to pinpoint organizational risks.
Michael Adams
Program Executive for AI Solutions,
Carahsoft

Why Deepfake Simulation Matters

Deepfake fraud has already cost organizations billions globally. Attackers are no longer limited to fake emails or static phishing pages. Instead, they’re cloning real voices, calling employees, and manipulating them into revealing sensitive information.
Traditional security training isn’t designed for this new reality. Slide decks and static phishing tests can’t replicate the stress of answering what sounds like your boss’s voice on the other end of a call.
That’s where simulation changes the game.
Resemble AI’s platform uses hyper-realistic voice cloning and adaptive AI to run live phishing scenarios. Imagine a spoofed customer call asking for credentials, or a WhatsApp message from a cloned executive voice. Employees are scored on their responses, while organizations get detailed analytics to highlight blind spots and areas of risk.

Why Carahsoft Partnership Matters

As Resemble AI’s Master Government Aggregator®, Carahsoft will make this platform widely accessible through its extensive contract vehicles, including NASA SEWP V, ITES-SW2, NASPO ValuePoint, and OMNIA Partners.
This means agencies can adopt Resemble AI’s platform quickly, without procurement hurdles, and begin strengthening defenses against deepfakes immediately.

Learn more about Resemble’s Deepfake Detection Platform

• Learn more about Resemble’s Deepfake Detection Platform
• Start a free trial

Original source

Two months ago, we did something the industry said was impossible. Our team released Chatterbox—an open-source text-to-speech model that outperformed ElevenLabs with 63.75% user preference in blind evaluations.
The response was overwhelming. Within weeks, Chatterbox exploded to over 1 Million downloads on Hugging Face and above 11,000 GitHub stars. Developers built audiobook generators, game studios integrated it for NPC dialogue, and educators created language learning tools. DigitalOcean called it “significant progress in personalized Voice AI.” The community had spoken: open-source TTS had finally arrived.
But in every GitHub issue, Reddit thread, and Discord conversation, one request dominated: “When will Chatterbox support other languages?”
Today, we answer that call.

Languages Supported

• Arabic (ar)
• Danish (da)
• German (de)
• Greek (el)
• English (en)
• Spanish (es)
• Finnish (fi)
• French (fr)
• Hebrew (he)
• Hindi (hi)
• Italian (it)
• Japanese (ja)
• Korean (ko)
• Malay (ms)
• Dutch (nl)
• Norwegian (no)
• Polish (pl)
• Portuguese (pt)
• Russian (ru)
• Swedish (sv)
• Swahili (sw)
• Turkish (tr)
• Chinese (zh)

State of the Art Voice Cloning in 23 Languages

Demand for multilingual TTS is skyrocketing. Brands, educators, and developers are no longer limited to English-only voice AI. Users expect apps and agents to sound human, speak in their native language, and deliver content with authentic tone and emotion.
Chatterbox Multilingual was built to meet that demand. It combines:

• Breadth of languages: 23 supported languages from launch.
• Expressive control: Fine-tune delivery with emotion and intensity settings.
• Enterprise reliability: Ultra-stable inference and built-in watermarking.
  Whether you’re designing a voice AI agent for customer support, a language-learning app, or a global gaming experience, Chatterbox Multilingual gives you the tools to build voices that resonate everywhere.

Chatterbox Multilingual Samples

Get Started from HuggingFace

Six lines of code to access Hollywood-quality voice synthesis in 23 languages.

Enterprise Level Voice Cloning with Chatterbox Multilingual Pro

While the open-source release brings world-class TTS to everyone, many enterprises need fine-tuned performance and strict service guarantees. That’s where Chatterbox Multilingual Pro comes in.

• 🎯 Custom fine-tuning: Train on your brand’s vocabulary, product names, and domain terms.
• ⚡ Sub-200 ms latency: Real-time streaming with global infrastructure.
• 📈 SLAs and reliability: Guaranteed uptime and throughput.
• 🔒 Advanced watermarking & detection: Enhanced protection against deepfakes.
  For call centers, financial services, healthcare platforms, and any organization where accuracy and compliance matter, Chatterbox Multilingual Pro closes the last mile.

Get started with Chatterbox Multilingual for free

Create your Voice Clone

Original source

Resemble AI is raising the bar for inline in-call detection with new support for leading telephony codecs — G.711, G.729, AMR-WB, and Opus — combined with a significant accuracy breakthrough in detecting synthetic and manipulated speech across compressed audio streams.

Why it matters

Most voice fraud today rides over bandwidth-constrained codecs optimized for telephony. Detecting deepfakes or impersonation attacks in these conditions is notoriously challenging due to lossy compression, packetization, and bandwidth artifacts. Our latest release closes this gap, ensuring high-precision detection even on low-bitrate, real-world call traffic.

What’s new

• Codec-Aware Detection Models – Enhanced accuracy across narrowband and wideband codecs without requiring PCM-quality input.
• Improved Inline Reliability – Greater resilience to compression artifacts, jitter, and packet loss in SIP, SIPREC, and RTP streams.
• Enterprise & Carrier Scale – Optimized for high concurrency deployments in contact centers, SBC environments, and telco networks.
• Security Against Evasion – Models hardened against codec-level adversarial perturbations, closing common attack vectors.

The outcome

Enterprises and carriers can now deploy real-time, in-call detection with confidence that voice integrity checks remain accurate regardless of the codec in use. Whether protecting a financial services hotline, a carrier call path, or a government comms channel, detection remains reliable, fast, and inline.

About Resemble’s Deepfake Detection Model

Resemble AI’s Deepfake Detection Model is a multimodal system built to identify manipulated audio, images, and video with industry-leading accuracy. Trained on a wide range of generative model outputs, it can spot synthetic media even as new architectures and techniques emerge. The model operates in real time, making it ideal for securing customer interactions, financial transactions, and enterprise communications where trust is critical. By combining audio forensics, visual signal analysis, and cross-modal consistency checks, Resemble delivers a unified layer of protection against deepfakes, ensuring authenticity across every channel where fraud may appear.

Ready to deploy realtime Deepfake Detection?

Schedule Free Demo

Original source

Today, Resemble AI is excited to introduce a groundbreaking approach to cybersecurity: a voice-based deepfake simulation platform designed to help organizations test and harden their defenses against AI-driven social engineering. Early adopters have already reported up to a 90% reduction in successful attacks.

While most security awareness training relies on static videos and generic phishing emails, Resemble AI’s platform brings simulations to life with hyper-realistic voice cloning and adaptive conversations. Think of it as red-teaming, but with AI-powered impersonation attempts sent straight to your employees — via phone calls, voicemails, and even WhatsApp messages.

These aren’t scripted prompts. Our platform uses agentic AI that maintains context and handles objections in real-time, powered by Resemble’s proprietary voice models and LLM integrations. From a spoofed CFO voicemail asking for wire transfers to a simulated customer demanding account access, employees experience real-world pressure, with none of the real-world fallout.

Every simulation generates a risk score from 0 to 100, personalized at the individual, team, and department level. Security teams can instantly identify blind spots and high-risk behavior. Whether it’s someone clicking a suspicious link or revealing sensitive info to a fake executive. The result? Targeted training where it’s actually needed, not across-the-board compliance theater.

This simulation platform is already in limited release, piloted by leaders in finance, healthcare, and customer service — sectors where deepfake-based fraud is both prevalent and devastating. In industries where frontline workers answer every call and make real-time decisions, deepfake protection can’t be theoretical. It has to be tested.

According to Resemble AI’s deepfake incident tracker, voice-based fraud has already caused over $2.6 billion in damages globally. Our new platform allows companies to take a proactive approach, actively simulating how those attacks would unfold—and giving them a chance to intervene before real money, data, or trust is lost.

“Today’s attackers aren’t playing by the old rules,” said Zohaib Ahmed, CEO and Co-founder of Resemble AI. “They’re using cloned voices, social cues, and urgency to manipulate in real time. We built this platform so companies can experience what that looks like inside their own walls—and build muscle memory before it counts.”

A Shift From Passive to Active Defense

Legacy tools like KnowBe4 and Cofense focus on passive learning. Resemble AI’s approach is active, adaptive, and voice-first. We don’t just test your weakest link, we help you reinforce it.

Recent deepfake scams, like the one targeting Accenture’s CEO in May, show how voice impersonation is becoming the go-to vector for attackers. Whether it’s a fake vendor, a spoofed HR exec, or a “friendly” internal request, these attacks often sound legitimate and slip past traditional training methods. Our simulations are designed to mimic that reality.

Built Ethically. Deployed Securely.

Resemble AI leads with ethical guardrails and secure design. Our platform integrates:

• Resemble Detect for deepfake detection
• Resemblyzer for speaker verification
• Resemble Watermark for forensic watermarking

We’re serious about safety.

Want to see how your team stacks up against voice-based deepfakes?

Request Access to the Simulation Platform

Original source

The deepfake arms race has created a deeply unfair fight. On one side: increasingly powerful AI creation tools with slick interfaces anyone can use. On the other: detection capabilities hidden behind enterprise paywalls, academic jargon, and technical complexity.
It’s like we’ve given everyone flamethrowers while locking fire extinguishers in special high-security vaults that require advanced degrees to access.
We’ve had enough of this asymmetry, so we built something that flips the script: a deepfake detection service that works through WhatsApp. Text 218-NO-FAKES, send any suspicious media, and get an immediate analysis that doesn’t just tell you if something is manipulated, but how it was likely created.

The Simplicity Imperative

When we started designing this service, we obsessed over one thing: removing friction. Every step, every requirement, every technical hurdle would mean fewer people using it.
So we opted for the most ubiquitous messaging platform on Earth. No downloads. No accounts. No subscriptions. No specialized knowledge required.
Just send media to 218-NO-FAKES.
Our detection technology works across modalities – analyzing images, video, and audio with 94% accuracy. But accuracy isn’t enough if the tools remain inaccessible.

Real-World Verification in Action

We’ve been testing this service across a spectrum of everyday scenarios where the line between authentic and synthetic content gets blurry. Let me walk you through a few examples:

The Digital Receipt Runaround

We took a ChatGPT-generated receipt – the kind of thing someone might use to fake an expense report or establish a false alibi – and sent it to our detector. Within seconds, it flagged the telltale patterns of AI generation. The subtle pixel-level irregularities that human eyes miss? Our system caught them immediately. No more wondering if that invoice is legitimate.

The Forwarded Fear-Mongering

We’ve all been there – someone forwards an alarming image on iMessage that looks just credible enough to cause anxiety. Is that really a hurricane heading toward your city? Did that politician actually do that embarrassing thing? Our system lets you forward that content directly from your messaging apps to 218-NO-FAKES, providing clarity when you need it most.

The “Wait, Was That Really My Bank?” Check

Perhaps most chilling is how easily voice synthesis can now mimic people we trust. We tested a synthetic voicemail that sounded convincingly like it came from a financial institution. Our detector immediately highlighted the voice as synthetic, pointing out the micro-patterns that distinguish AI-generated speech from human voices.

Join the Verification Movement

Text 218-NO-FAKES on WhatsApp and try it yourself. Send that suspicious image. Forward that weird audio clip. Check that too-perfect video.
The more people with verification tools, the less incentive there is to create malicious deepfakes in the first place. It’s a simple equation: when deception becomes pointless, it stops being profitable.
This is just the beginning. We’re already working on enhanced detection capabilities, additional platforms, and deeper educational resources.
Because in a world increasingly shaped by artificial intelligence, the power to discern truth shouldn’t be artificial, too.
It should belong to all of us.

Original source

The Power of Description

Today, we’re thrilled to unveil Voice Design, our most groundbreaking feature yet. Voice Design represents a fundamental shift in how creators approach voice generation by translating simple text descriptions into fully-realized AI voices in seconds.

Creating the perfect voice has traditionally been a complex process involving multiple steps, technical parameters, and often, compromise. Voice Design changes that paradigm entirely.
With Voice Design, you simply describe the voice you want in natural language—as if you were explaining it to a colleague:

• “An old school pirate, speaking with a hearty growl, stretching vowels dramatically”
• “A calm podcast host with a deep, resonant voice and slight southern accent”
• “Energetic TV game show host with rapid delivery and enthusiastic tone”
  Our advanced AI then analyzes this description, identifying key voice characteristics, tonal qualities, speech patterns, and stylistic elements to generate a voice that matches your imagination.

Unlimited Creative Possibilities

Voice Design isn’t just a tool—it’s a canvas for voice creativity without boundaries. Every description produces a unique voice, meaning your creative options are virtually limitless. By harnessing the power of natural language descriptions, we’ve transformed voice creation from a technical process to an intuitive, creative act accessible to everyone.
The core innovation of Voice Design lies in its ability to interpret nuanced descriptions and translate them into distinctive vocal characteristics—from subtle inflections to dramatic personality traits. Whether you need a voice that conveys authority, warmth, excitement, or any combination of qualities, it’s now just a description away.

Transforming Industries Through Voice Innovation

The applications of Voice Design extend far beyond conventional use cases, revolutionizing workflows across numerous industries:

Content Creators

For animators, podcasters, and video producers, Voice Design revolutionizes the creative process by offering unprecedented flexibility in voice creation. Content creators can now experiment with different narrator styles for documentaries, develop distinct character voices for animations, and establish unique vocal identities for channels or series—all without the scheduling and budget constraints of traditional voice recording. This freedom to iterate quickly means creators can explore creative directions that would have been impractical before, ultimately leading to more distinctive and engaging content that resonates with audiences.

Game Developers

Game development teams face unique challenges when creating immersive worlds populated by dozens or even hundreds of characters. Voice Design transforms this aspect of game creation by enabling developers to rapidly prototype character voices during early development, test how different vocal styles affect emotional impact, and create distinctive voices for non-player characters throughout the game world. This capability not only streamlines the development process but also opens up new possibilities for creating richer, more diverse game environments where each character can have a truly unique vocal identity.

Educational Content

In educational settings, engagement is everything. Voice Design enables educators and e-learning developers to create varied narrator voices across different modules, keeping students engaged through long learning sessions. By tailoring voices to specific subject matter, age groups, or learning contexts, educational content becomes more accessible and compelling. Language learning applications particularly benefit from the ability to generate consistent pronunciation examples, while complex subjects can be explained in voices specifically designed to emphasize clarity and comprehension, ultimately enhancing learning outcomes.

Accessibility Solutions

Voice Design represents a significant advancement in creating truly inclusive audio experiences for diverse audiences. Developers can now design voices optimized for clarity and comprehension for visually impaired users, create voices that speak at adjustable paces for people with cognitive processing differences, and develop region-specific accents to make content more relatable across global markets. This level of customization means that accessibility is no longer about finding the closest acceptable option, but rather creating voices specifically designed to meet the unique needs of different user groups.

Advertising and Marketing

Brand voice has always been a critical but abstract concept. With Voice Design, marketers can literally create voices that embody brand personalities, turning conceptual brand attributes into audible experiences that resonate with audiences. Agencies can quickly prototype different vocal approaches for client presentations, ensure consistency across campaign elements, and develop seasonal variations that maintain brand recognition while conveying different emotional tones. This capability transforms how brands think about their audio identity, making voice a more central and strategic element of the marketing mix.

Entertainment and Performance Arts

The entertainment industry thrives on innovation, and Voice Design offers creators new tools for pushing creative boundaries. Directors and producers can conceptualize character voices before casting decisions, generate atmospheric background voices for immersive scenes, and experiment with unconventional vocal styles that might be difficult to achieve through traditional means. This technology doesn’t replace performers, but rather expands the creative palette available to artists, allowing them to explore new possibilities and refine their vision before moving into production.

Voice Design in Action

Here are some examples showcasing the versatility of Voice Design:

Example 1: Character Voices

Wise elderly wizard with a slightly raspy voice, speaking slowly and deliberately.

Example 2: Narrator Styles

A super deep male voice announcer with a dramatic performance that is unveiling Resemble AI's new Voice Design feature

Example 3: Creative Concepts

Robot assistant with a friendly, slightly mechanical voice that speaks with perfect diction

The Bigger Vision

Beyond these specific features, our long-term vision is to democratize voice creation completely. We believe that voice, like visual design, should be a creative medium accessible to everyone regardless of technical expertise.
We envision Voice Design becoming an essential tool in creative workflows—where generating the perfect voice is as simple as describing what you hear in your imagination. This vision extends to creating a marketplace where voice creators and consumers can collaborate, opening new possibilities for customization and personalization.
As AI voice technology evolves, we remain committed to responsible development—prioritizing transparency, quality, and creative empowerment while building tools that amplify human creativity rather than replace it.
We invite you to join us on this journey, share your feedback, and help shape the future of voice creation together.

Try Voice Design

Original source