What is AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent)?
AssemblyAI-STT-3.7 is a family of three speech-to-text models announced on 18 November 2026. STT-3.7 handles recorded audio; STT-3.7-Realtime is for live meetings, calls and captions, with speaker labels in the stream; and STT-3.7-Realtime-Agent is tuned for voice agents, listening to the person talking to the bot and deciding when a turn has ended.
All three cover 32 languages with automatic detection and switching between languages mid-conversation, up from 18. For recorded audio, adding the older Universal-2 model extends coverage to 99 languages. Users can steer vocabulary with key terms and, on the real-time models, a 400-character prompt.
The family replaces Universal-3.5 Pro and the Universal-Streaming models. Endpoints and message formats are unchanged, but two defaults flip: filler words are now kept and text formatting is on.
How much does AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) cost?
| Plan | Price | What you get |
|---|---|---|
| STT-3.7 | $0.21 an hour | Recorded audio, 32 languages, Universal-2 fallback to 99 |
| STT-3.7-Realtime | $0.30 an hour | Live meetings, calls and captions with live speaker labels |
| STT-3.7-Realtime-Agent | $0.45 an hour | Voice agents, with end-of-turn detection |
STT-3.7 $0.21 an hour; STT-3.7-Realtime $0.30 an hour; STT-3.7-Realtime-Agent $0.45 an hour. Speaker labels +$0.02 an hour on recorded audio. Universal-2 stays at $0.15 an hour; $50 free credit.
Is AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) worth it?
AssemblyAI built this release around the phone call, and it shows. On its own set of 12,460 voice-agent calls across 26 streaming models, STT-3.7-Realtime-Agent made 4.53% word errors, against 6.02% for Soniox's new v6, and cut number errors from 11.28% to 2.27% compared with its predecessor. The final transcript arrives 86 milliseconds after the caller stops speaking. The independent Voice Arena board also puts it first for US English streaming, at 2.45% semantic error.
Recorded audio is a closer race. Artificial Analysis measures STT-3.7 at 2.5% word error rate, 12th on its board and behind ElevenLabs Scribe v3 at 1.9%. Where AssemblyAI does lead is speaker labelling: 10.7% diarization error across six English conversation sets, against 12.0% for ElevenLabs, with no other provider it tested under 15%. Those are AssemblyAI's figures, run between August and November.
The catch: the defaults change under you. Filler words and formatting now come back on by default, so downstream code that expects clean, unpunctuated text needs a check, and recorded audio outside the 32 languages no longer falls back automatically unless you list Universal-2. Universal-3.5 Pro Realtime retires on 1 February 2027.
Our take: for a voice agent, $0.45 an hour buys the model with the most evidence behind it on noisy, number-heavy calls. For a transcript archive, the batch model at $0.21 is a sound default, but buyers chasing the lowest error rate will find two or three rivals ahead of it.
Where it earns its place
A pharmacy chain's automated refill line
Catch prescription numbers and dates of birth on the first attempt
Stream calls to assemblyai-stt-3-7-realtime-agent with key terms for drug names. AssemblyAI reports number errors down 80% on its predecessor; 10,000 call hours cost about $4,500.
A sales-analytics start-up
Label who said what on recorded discovery calls in English, Spanish and German
Batch STT-3.7 with speaker labels at $0.23 an hour ($0.21 plus $0.02 for diarization); mixed-language calls are handled without choosing a language first.
Before you commit
- Setup
- Set speech_models to ["assemblyai-stt-3-7"] on the existing transcript endpoint, or connect the streaming socket with assemblyai-stt-3-7-realtime or -realtime-agent. No protocol changes.
- Integrations
- AssemblyAI REST and streaming APIs and SDKs; reference agents for LiveKit and Pipecat; pairs with AssemblyAI-TTS-1 for end-to-end turns of about 250 ms in AssemblyAI's tests.
- Data and privacy
- AssemblyAI's data policy trains on customer audio by default after redacting personal information; paid accounts can opt out, and BAA and EU accounts are always opted out. No published weights.
- Cost at scale
- 100,000 hours of recorded audio cost about $21,000; the same volume on the agent model is $45,000. Streaming is billed for the whole time a socket is open, idle time included.
- Limits
- 32 native languages; others need Universal-2. Real-time model has no turn detection (use the Agent model). Universal-3.5 Pro Realtime retires 1 February 2027.
What’s good
- 4.53% error on 12,460 agent calls, lowest of 26 streaming models in AssemblyAI's test
- Final text 86 ms after speech ends on the agent model
- 10.7% diarization error on recorded audio, ahead of ElevenLabs
- 32 languages with code-switching; 99 with Universal-2
What’s not
- 2.5% on Artificial Analysis, 12th and behind Scribe v3
- Most comparison figures are AssemblyAI's own
- Changed defaults can break parsers built for Universal-3.5 Pro
- Agent model costs more than twice the batch rate
Best for
Teams building phone or web voice agents who need reliable numbers, names and turn-taking, and call-analytics teams who care most about who said what.
AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) alternatives
AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent): common questions
- Is AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) worth it?
- The voice-agent model is where AssemblyAI earns its place; on plain recorded audio it is good, cheap and no longer quite the best. We rate it 4.4 out of 5.
- How much does AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) cost?
- STT-3.7 $0.21 an hour; STT-3.7-Realtime $0.30 an hour; STT-3.7-Realtime-Agent $0.45 an hour. Speaker labels +$0.02 an hour on recorded audio. Universal-2 stays at $0.15 an hour; $50 free credit.
- What are the downsides of AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent)?
- 2.5% on Artificial Analysis, 12th and behind Scribe v3; most comparison figures are AssemblyAI's own; changed defaults can break parsers built for Universal-3.5 Pro; agent model costs more than twice the batch rate.
- Who is AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) best for?
- Teams building phone or web voice agents who need reliable numbers, names and turn-taking, and call-analytics teams who care most about who said what.
- How much does AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) cost for a team or in production?
- 100,000 hours of recorded audio cost about $21,000; the same volume on the agent model is $45,000. Streaming is billed for the whole time a socket is open, idle time included.
- Is AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) safe for company data?
- AssemblyAI's data policy trains on customer audio by default after redacting personal information; paid accounts can opt out, and BAA and EU accounts are always opted out. No published weights.
- What are the best AssemblyAI STT-3.7 (batch, Realtime and Realtime-Agent) alternatives?
- The closest alternatives we have reviewed are ElevenLabs Scribe v3 and Scribe v3 Live, Soniox STT v6 (stt-rt-v6 and stt-async-v6), Deepgram Nova-3 and Voxtral 3 (Transcribe and TTS).
Answers checked Nov 21, 2026.