AI NewsModels & agentsAnnouncement
AssemblyAI released STT 3.7 with 32 languages and a diarization model
AssemblyAI released its STT 3.7 family on 17 November with three builds, 32 languages and a new diarization model.

Image: AssemblyAI
Why it mattersA team on an older transcription API gets more languages and a cleaner speaker split for the same hourly price. The gain is on the vendor's own tests.
A team paying for a speech-to-text API got a new model family from AssemblyAI. The company released STT 3.7 on 17 November with three builds and a wider language list.
AssemblyAI posted the announcement on its own blog that day. Tom Shapland wrote it. The main build is STT 3.7 for recorded audio. The live-call build is STT 3.7 Realtime. The voice-agent build is STT 3.7 Realtime Agent. All three share one model family and are generally available today.
What is new
AssemblyAI says the family covers 32 languages, up from 18. The company says the model handles native code switching. A call that moves between two languages mid-sentence is handled inside one request.
A separate research post by Britney Xiu names the headline number on speaker labelling. The new diarization model has a 29 percent lower Diarization Error Rate than the previous one, in AssemblyAI's own tests. Diarization is the step that says who spoke each word. A lower rate means fewer wrong speaker labels on a multi-speaker call.
The pricing line did not change. AssemblyAI lists the family at $0.21 per hour, the same rate as the previous version.
Who the three builds are for
The recorded-audio model is for files already on disk. The full audio is available when the model reads it. The real-time model is for a live stream of audio with a low delay. The voice-agent model is tuned for a two-way agent that speaks back. AssemblyAI pairs it with its text-to-speech API and its Speech Understanding API. Those run at the same time on the same audio.
What AssemblyAI does not publish
The blog does not share a Word Error Rate for the main transcription model. It does not compare STT 3.7 to a competitor on the same dataset. It does not list a latency figure for the real-time model. The research post names the 29 percent gain on the vendor's own diarization test set only.
The company names two customers on the page. Zoom is listed as a user of AssemblyAI technology. Siro is quoted with a 90 percent reduction in customer complaints and support tickets, which is Siro's own number. Neither is tied to the STT 3.7 release itself.
What a developer should do today
A team already on AssemblyAI can switch the model name in its client and keep the same contract. A team on another provider gets a short list of comparisons to run on real audio. Pick the 100 worst calls from the last month. Pick the ones with cross-talk, accents, or background noise. Run the real-time build on the same clips through both providers. Count the wrong speaker labels by hand on 20 of them. The 29 percent figure is AssemblyAI's own, on its own previous model. Your number on your audio is the one that decides the switch.
Source
AssemblyAI blog: Introducing AssemblyAI-STT-3.7. Technical details in the research post by Britney Xiu.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

