Tasks and inputs
AssemblyAI-STT-3.7-Realtime-Agent is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 18, 2026
AssemblyAI-STT-3.7-Realtime-Agent is a streaming recogniser for voice agents and IVR: 32 languages, 86 ms to final transcript, 4.46% vendor WER, $0.45/hour.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Indicative: $0.45 per audio hour. Confirm on your AssemblyAI account.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Numbers spoken over a phone line break voice agents more often than accents do. A caller reads a policy number, the recogniser drops a digit, and the whole flow fails. AssemblyAI's voice-agent model for November 2026 cut number errors by about 80% on its own test, from 11.28% to 2.27%.
AssemblyAI-STT-3.7-Realtime-Agent (assemblyai-stt-3-7-realtime-agent) replaces Universal-3.6 Pro Realtime at the same $0.45 per hour. It launched on 17 November 2026.
Against its predecessor, on AssemblyAI's own voice-agent set:
| Metric | Before | After |
|---|---|---|
| Word error rate | 5.19% | 4.46% |
| Number errors | 11.28% | 2.27% |
| Entity misses | 23.28% | 11.00% |
| Wrong short replies | 1.35% | 0.96% |
| Background words per 1,000 | 2.53 | 1.34 |
| Partial-to-final word changes | 3.2% | 1.0% |
False end-of-turn events fell 12%. A final transcript now lands 86 ms after speech ends, down from 172 ms, while end-of-turn latency itself is unchanged.
Background speech, words picked up while the caller is silent, was roughly halved but not eliminated. If your callers sit in noisy offices, keep the Voice Focus feature on.
On a public voice-agent benchmark published on 16 November 2026, the model ranked second on the Pipecat WER set (up from fourth), seventh on overall WER and tenth on another conversational set. Time to final transcript there was 47 ms. Time to first partial, at 493 ms, slipped from first to fourth place. So it is not the fastest to show early text, and AssemblyAI's own guidance stops short of claiming the lowest background-speech rate, since one competitor does better on that.
Turn detection, keyterms, prompting, Medical Mode and Voice Focus all work as before. The endpoint and protocol are unchanged; migration is a change of speech_model. Language coverage is the same 32 as the other 3.7 variants, with automatic detection.
Replay 200 recorded calls with known outcomes through your agent stack twice, once per model. Track task completion, digit-string accuracy, false barge-ins and the time from end of speech to agent response. If you do not need turn-taking at all, the cheaper Realtime model is a better fit. A speech-to-speech alternative worth including is GPT-Realtime-2.1, and more streaming recognisers are listed under speech to text. All figures are vendor-reported and prices indicative.
Compare the model on the work you need to complete.
AssemblyAI-STT-3.7-Realtime-Agent is listed for Automatic speech recognition, Speech → Text. It accepts audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
Use the model identifier and account requirements for the selected API route. Confirm its request limits and output format with one complete task before connecting a larger workflow.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
assemblyai
AssemblyAI-STT-3.7 is AssemblyAI's November 2026 pre-recorded transcription default: 32 languages, new diarization, fewer language flips, $0.21 per hour.
assemblyai
AssemblyAI-STT-3.7-Realtime streams transcripts in 32 languages for captions, notetakers and agent assist, with final speaker labels per turn at $0.30/hour.
salesforce
SOV-Transcribe-1 is Salesforce's open-weight speech-augmented model for non-streaming transcription, with keyword biasing and style instructions; Apache-2.0.
ElevenLabs
Scribe v3 Live is ElevenLabs' streaming transcription model with turn prediction, mid-sentence language switching and a 10,000-term keyterm library.
NVIDIA
Nemotron 3.5 Transcribe 30B A4B is NVIDIA's open offline ASR model: 40 languages, hour-long files, word timestamps, diarisation and translation.
NVIDIA
Nemotron-Labs-Unified-ASR-0.6b is NVIDIA's open 600M speech recognizer for 25 European languages, running offline or streaming at down to 160 ms.