# Is AssemblyAI-STT-3.7-Realtime-Agent SOTA?

Pending benchmarks. Announced Nov 17, 2026; no independent results yet.

As of Nov 22, 2026, AssemblyAI-STT-3.7-Realtime-Agent is pending benchmarks on isitsota: announced Nov 17, 2026, with no independent results on the boards we track yet.

AssemblyAI's streaming speech-to-text model in the AssemblyAI-STT-3.7 family, built for voice agents. Official docs describe 32 languages in one stream with automatic detection, plus keyterm prompting, context carryover, conversation memory, and Voice Focus. It replaces Universal-3.6 Pro Realtime for voice-agent use. Live on AssemblyAI's Realtime Speech-to-Text API as speech_model="assemblyai-stt-3-7-realtime-agent" at $0.45 per hour of session time. No weights. Official models.md dates the family launch to November 17, 2026; the changelog and lab blogs are dated November 18.

Canonical: https://isitsota.com/m/assemblyai-stt-3-7-realtime-agent
Last updated: 2026-11-18
Tracking since: 2026-11-18

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | AssemblyAI |
| Announced | Nov 17, 2026 |
| API | Live (`assemblyai-stt-3-7-realtime-agent`) · Realtime Speech-to-Text API: WebSocket wss://streaming.assemblyai.com/v3/ws with speech_model="assemblyai-stt-3-7-realtime-agent" and an API key. Official introducing post: the endpoint and message protocol are unchanged from Universal-3.6 Pro Realtime |
| Open weights | Closed · Not released |
| Task | Streaming speech-to-text for voice agents, including multilingual agents with native code switching and agents in noisy environments |
| Languages | 32 languages in one stream, with automatic detection and native code switching (official introducing post, models reference, and which-model guide) |
| Pricing | $0.45 per hour of session time, the same rate AssemblyAI states for Universal-3.6 Pro Realtime (official pricing reference and Realtime-Agent research post). Official Voice Focus docs: Voice Focus is available on this model as near-field or far-field, at +$0.10/hr |
| Features | Official streaming docs: built-in keyterm prompting, context carryover, and conversation memory, and Voice Focus (near-field or far-field). Official research post: the model receives text context and acoustic context from earlier in the call; it is trained on real voice-agent and telephony conversations with speaker and role labels |
| Availability | Generally available on AssemblyAI's self-serve Realtime Speech-to-Text API. Official changelog dated 2026-11-18. Official docs list it as the replacement for Universal-3.6 Pro Realtime in voice-agent use |
| Related | Separate from AssemblyAI-STT-3.7 (pre-recorded) and from AssemblyAI-STT-3.7-Realtime (live conversations and notetakers), from AssemblyAI-TTS-1, and from Universal-3.5 Pro |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim: on 12,460 English voice-agent calls, AssemblyAI reports a lower word error rate and a lower entity error rate than the other 25 streaming models it ran, including Soniox Realtime v6, ElevenLabs Scribe v3 Live, and Deepgram Nova-3 (self-reported, official Realtime-Agent research post).
- Lab claim: it returns the final transcript 86 ms after the caller stops speaking, half the 172 ms AssemblyAI measured for Universal-3.6 Pro Realtime on the same day (self-reported, official research post).
- Lab claim: versus Universal-3.6 Pro Realtime, end-of-turn F1 rose from 83.60% to 85.52%, precision from 76.08% to 78.97%, and false end-of-turn events fell from 23.92% to 21.03% (self-reported, official research post).

## Sources

- LAB: [AssemblyAI models reference (llms/models.md)](https://www.assemblyai.com/llms/models.md) (2026-11-17)
- LAB: [Speech model versions and deprecations](https://www.assemblyai.com/docs/getting-started/model-versions-and-deprecations) (2026-11-17)
- LAB: [AssemblyAI changelog: AssemblyAI-STT-3.7](https://www.assemblyai.com/changelog/) (2026-11-18)
- LAB: [Introducing AssemblyAI-STT-3.7 (AssemblyAI blog)](https://www.assemblyai.com/blog/introducing-assemblyai-stt-3-7) (2026-11-18)
- LAB: [AssemblyAI-STT-3.7-Realtime-Agent research post](https://www.assemblyai.com/blog/assemblyai-stt-3-7-realtime-agent-research) (2026-11-18)
- LAB: [Streaming model selection](https://www.assemblyai.com/docs/streaming/select-the-speech-model)
- LAB: [Voice Focus](https://www.assemblyai.com/docs/streaming/voice-focus)
- LAB: [AssemblyAI pricing reference](https://www.assemblyai.com/pricing.md)
- LAB: [AssemblyAI on X: AssemblyAI-STT-3.7](https://x.com/AssemblyAI/status/2123080177508929849) (2026-11-18)

Tracking continues until an independent board lists it.

## Related

- Lab: [AssemblyAI](https://isitsota.com/labs/assemblyai.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/assemblyai-stt-3-7-realtime-agent
