# New on AssemblyAI: New STT and TTS models, LLMs and more.

Canonical: https://brew.new/templates/assemblyai/new-on-assemblyai-new-stt-and-tts-models-llms-and-more

Brand: assemblyai.com
Category: re-engagement

![Preview of New on AssemblyAI: New STT and TTS models, LLMs and more.](https://cdn.brew.new/email-preview-9d444f6546a335e7-tracking_r57gcq7bcxmfe9s794myyxnn618jna4b-1795227659993.png)

## Email content

AssemblyAI Newsletter

Hey there,

TL;DR: AssemblyAI-STT-3.7 is here, with three transcription models for live conversations, voice agents, and recordings. AssemblyAI-TTS-1 brings text-to-speech built for voice agents, and there’s a faster Sync API, 20-minute Dictation API requests, an official MCP server, and new models on the LLM Gateway.

New models

AssemblyAI-STT-3.7: a transcription model for every kind of audio

AssemblyAI-STT-3.7 has three transcription models, each built for one kind of audio: live conversations, voice agents, and recordings. All three now cover 32 languages, up from 18, with automatic detection.

Introducing AssemblyAI-STT-3.7, a family of models Live speaker labels for meetings and calls. AssemblyAI-STT-3.7-Realtime posts a 6.30% live diarization error rate on meeting audio, ahead of ElevenLabs Scribe v3 Live at 9.63%.

Built for voice agents. AssemblyAI-STT-3.7-Realtime-Agent scores a 4.53% word error rate on 12,460 voice agent calls, ahead of ElevenLabs, Deepgram, Soniox, and more.

Recordings with fewer speaker errors. AssemblyAI-STT-3.7 posts a 10.7% diarization error rate across six English datasets, 36% fewer than the competitive average.

Read the announcement

Introducing AssemblyAI-TTS-1

AssemblyAI-TTS-1 is our first text-to-speech model, built for live calls. It starts speaking in about 90 ms, runs with unlimited concurrency and no annual contract, and costs $15 per 1M characters. Paired with AssemblyAI-STT-3.7-Realtime-Agent and the LLM Gateway, full turns regularly land around 250 ms.

More than 4,000 voices, or your own. Pick from the library, or clone a voice instantly from a short recording. Cloned voices cost the same $15 per 1M characters.

Six languages, in the US or the EU. English, Spanish, French, German, Italian, and Portuguese, with accents, and more on the way. EU processing costs the same as US.

Ready for LiveKit and Pipecat. Reference agents for both are on GitHub. The stream returns word timestamps and accepts a cancel message, so barge-in works without extra plumbing.

Try it in the playground

Releases & updates

Other product updates this month

The Sync API is faster and cheaper: responses now come back in about half the time, with no changes to your code, and the list price is now $0.30/hr, down from $0.45/hr. Read the Sync API guide

The Dictation API takes 20-minute recordings: send a finished recording of up to 20 minutes to POST /v1/transcribe, up from 2 minutes, and set rewrite_mode to fast or smart to choose speed or cleanup quality. See the docs

An official AssemblyAI MCP server: connect Claude, Cursor, VS Code, or any MCP client and transcribe, summarize, and analyze audio on your account without writing glue code. Set it up

New models on the LLM Gateway: DeepSeek V4.1 Flash 1017, Beam 1.0, and Muse Spark Open, with a 1M-token window on Muse Spark Open. LLM Gateway has 50+ models available with a single API. See the models

Deprecation · February 1, 2027

The Voice Agent API retires on February 1, 2027. To keep an agent running, stream audio to AssemblyAI-STT-3.7-Realtime-Agent and build the rest of the loop on LiveKit or Pipecat. Read the notice →

AssemblyAI Our speech-to-text models lead the industry in accuracy and deliver the most reliable audio outputs for standout user experiences.

169 Madison Ave

Suite 38365

New York, NY 10016 Unsubscribe

[Open and remix this design](https://brew.new/templates/assemblyai/new-on-assemblyai-new-stt-and-tts-models-llms-and-more)

[Browse email designs](https://brew.new/browse/templates)
