Resemble AI is a company, based in Toronto and San Francisco according to a December 2025 report, that sells synthetic voice generation and deepfake detection, and it publishes an open-source speech model called Chatterbox under the MIT license. No independent, published ranking of Resemble AI against rival products could be found, so a fair comparison has to rest on criteria and on shared public test sets rather than on any vendor's own accuracy figures.
Key takeaways
- Resemble AI's core products are voice synthesis and audio deepfake detection, not general image recognition.
- Its accuracy figures (up to 98 percent in 2023, 98 percent across 38 languages in 2025) are company claims, and a 2023 report said the rate fell to 87 percent on unfamiliar voices.
- The ASVspoof 5 challenge from 2024 is a public benchmark where attacks significantly compromised baseline detectors.
- Any vendor comparison should test on a buyer's own audio, with stated error rates for both missed fakes and false alarms.
What the company makes
Resemble AI began as a voice cloning and text-to-speech provider. Sources disagree on the founding year, giving 2018 in one report and 2019 in another, and the founders first showed the technology in 2019. On July 12, 2023, the company announced a USD 8 million Series A round led by Javelin Venture Partners with Comcast Ventures, and launched a tool called Resemble Detect that flags synthetic audio. On December 8, 2025, it announced a USD 13 million round that brought its reported total funding to USD 25 million, with Google's AI Futures Fund, Okta Ventures, and Sony's innovation fund among the investors.
The December 2025 report describes a shift from generating voices toward security. Its newer detector, Detect-3B Omni, is described as a model with three billion parameters that analyzes audio and video during live calls without needing a prior sample of the person's voice or face. That is the only place where the company's products touch images at all, and the old page title, which mentioned image and audio recognition, is closer to a description of the product category than of this one company.
Chatterbox: the open-source side
Resemble AI also maintains Chatterbox, a family of text-to-speech models released under the MIT license on GitHub. The original model has about 500 million parameters and an English-only focus. A multilingual version covers 23 languages, and smaller variants of 350 million and 110 million parameters target faster or on-device use. The repository states that every generated audio file carries an inaudible neural watermark called PerTh, and it includes a script for checking whether a clip contains that mark. The page also asks users not to apply the model to harmful purposes.
A watermark helps only when the checker is available and the mark survives editing. The repository claims it survives MP3 compression, which is a vendor statement. Audio produced by other tools carries no such mark, so a missing watermark proves nothing about whether a clip is real.
What published tests say about detectors
The best public yardstick is the ASVspoof series, a research challenge on spoofed speech. The fifth edition, run in 2024 and described in an overview paper on arXiv, used crowdsourced recordings rather than studio-quality data, and added adversarial attacks for the first time. The follow-up database paper reports about 2,000 speakers, compared with about 100 in earlier editions, and 32 different attack algorithms. The overview reports that attacks significantly compromised the baseline systems, while submitted systems produced substantial improvements.
Two lessons follow for anyone reading a detector's accuracy claim. A single percentage depends on which fakes were in the test, and a detector trained on known generators can lose ground on a new one. The 2023 coverage of Resemble Detect said as much about itself: up to 98 percent in the best case and 87 percent for voices or tracks it had not met before.
A checklist for comparing voice and deepfake tools
The table below is an illustrative framework, not a ranking, and no product is scored in it.
| Question | Why it matters | What to ask a vendor |
|---|---|---|
| Which fakes were in the test? | Detectors degrade on generators they did not see during training | A list of generators and the date of the test |
| What are the error rates? | Missed fakes and false alarms trade off against each other | Both rates at the same threshold, not one accuracy number |
| Is the test independent? | Self-reported results favor the reporter | A third-party evaluation or a public benchmark |
| What audio conditions apply? | Phone compression and noise change results | Results on telephone-quality audio |
| Where does the audio go? | Voice recordings are personal data | Storage location, retention, and reuse for training |
| How is consent handled? | Cloning a voice without permission creates legal exposure | Consent process for any voice sample |
Where public policy stands
The US Federal Trade Commission ran a Voice Cloning Challenge and, on April 8, 2024, announced its winners. The three top entries, AI Detect, DeFake, and OriginStory, shared USD 35,000, and Pindrop Security received a recognition award. The entries used different approaches: detecting synthetic speech, making a voice harder to clone, and proving that a recording came from a human at the moment of capture. Detection is therefore one of several defenses, and each has gaps.
This article gives general information about a company and a testing method. It does not recommend a product, and it is not legal or security advice. Organizations weighing a purchase should ask vendors for test evidence and consult qualified advisers.
FAQ
Is Resemble AI an image recognition tool?
Its main products are voice generation and deepfake detection. The December 2025 report says its newest detector also handles video, but image recognition in the general sense is not what the company is known for.
Is Resemble AI free?
The Chatterbox models are free to use under the MIT license. The company also sells hosted services, and prices change, so current terms must be checked with the company.
Can a detector prove that audio is fake?
No. A detector estimates a probability, and its results depend on the generators and recording conditions in its training and test data. Findings are best treated as one input alongside call-back checks and other context.
Are vendor accuracy numbers reliable?
They describe a specific test run by the vendor. Public benchmarks such as ASVspoof 5 and independent evaluations give a more comparable picture.
Related reading
- Limits of AI emotion recognition
- How pattern recognition works in AI
- AI compared with automated intelligence
- AI and deep learning compared
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Resemble AI hauls in USD 13M with a different approach to deepfake detection, SiliconANGLE, December 8, 2025
- Resemble AI raises USD 8 million and launches deepfake voice detector, July 12, 2023
- Chatterbox repository and model card, Resemble AI on GitHub and Hugging Face, 2026
- ASVspoof 5: Crowdsourced Speech Data, Deepfakes, and Adversarial Attacks at Scale, arXiv 2408.08739, 2024
- ASVspoof 5: Design, Collection and Validation of Resources, arXiv 2502.08857, 2025
- FTC Announces Winners of its Voice Cloning Challenge, Federal Trade Commission, April 8, 2024