Resemble AI

Today it is mainly a detector. Resemble’s about page says the company started in 2019 to solve a voice production problem in gaming and then shifted its focus to detection. It now says it is focused entirely on multimodal AI detection. The homepage and the pricing page agree: they sell detection for audio, video and images, plus Intelligence, Meetings, Identity, Watermarker and Agent Detection, and they show no voice generation product.

Voice generation has not vanished. The developer docs still list text-to-speech, speech-to-speech and voice cloning under Voice APIs, the changelog shows voice updates through June 2026, and a text-to-speech product page describes the open-source Chatterbox model. The two descriptions sit side by side, so the first decision is which product you want.

Detection is billed per second of audio or video and per image. The Flex plan has no subscription fee: you load credits, they do not expire, and you pay only for what you run. Team and Business cost a monthly fee and cut the detection rates by more than half.

The page shows Team at $350 a month billed monthly or $280 a month billed annually, and Business at $1,000 a month billed monthly or $800 a month billed annually. Team adds large and batch uploads and 5 seats; Business adds Single Sign-On and 20 seats.

Only for steady volume. Editorial assumption using the published rates: audio detection saves $0.02 a second on Team ($0.035 minus $0.015). The $350 monthly fee divided by $0.02 is 17,500 seconds, about 292 minutes or 4.9 hours of audio a month. A 10-minute call is 600 seconds, which adds $21 of Flex usage or $9 of Team usage. Team also has its $350 monthly subscription: if this is the only call that month, the totals are $21 and $359 respectively, before tax.

• Below 17,500 seconds of audio a month, Flex has the lower combined subscription-plus-usage cost under monthly billing.
• Above that threshold, Team has the lower combined cost in this audio-only calculation. Its per-second usage rate is lower at any volume, but that alone does not repay the subscription.
• Video saves $0.04 a second, so a video-heavy workload reaches break-even faster.

These are provider rates, not measured results, and they ignore taxes, seats and any volume discount from sales.

Treat them as the vendor’s claims. The homepage states accuracy of 99.5% for audio, 98% for video and 96% for images for its DETECT-World model, and a benchmark panel on the same page shows 95.8% for images and 98.2% for video. The two panels do not match, and a footnote says image and video results come from internal testing with external testing in progress. We did not test the detector.

The vendor is also cautious about itself: its ethics page says no detector is perfect and that results inform a decision your team makes. A June 22, 2026 changelog entry reports a new audio model with an equal error rate down to 4%, which is a different measure from accuracy. Run the detector on your own files, including recordings you are authorized to upload, before using a result to reject a caller or a claim.

The facts we could confirm are narrow. Resemble Ultra is the current text-to-speech model, and the model-versions page lists Chatterbox, Chatterbox-Turbo and every earlier model as end of life on the hosted API. The official Chatterbox repository carries an MIT license with zero-shot voice cloning, so a team can self-host it without the hosted API.

Cloning needs only about 10 seconds of audio according to the docs, but the same page says the Voice Cloning API requires a Business plan or higher, and the text-to-speech page marks WebSocket streaming as Business and above. The May 8, 2026 changelog says the first clone is free and each extra clone is $2. No page we read lists a price per character or per minute of generated speech, so the cost of generating audio at volume is unknown until you ask the vendor.

1. Collect 20 to 30 files you already know the truth about: non-sensitive recordings you own or have explicit permission to upload, and synthetic samples you made yourself.
2. Run them on Flex and note the cost per file against the rates above.
3. Record every false alarm on genuine files; those are the cases that hurt a real customer.
4. If you need voice generation, ask sales for the text-to-speech rate in writing before you build on it.
5. Save the pricing page and the terms version you relied on.