THE DIRECT ANSWER

AssemblyAI’s new STT 3.7 family offers separate models for recorded audio, live transcripts and conversational agents. The release expands language coverage, but the immediate practical choices are which model fits the application and whether existing requests will change automatically.

Read the dated primary source ↓

One release, three listening jobs

AssemblyAI has introduced its STT 3.7 speech-to-text family, with the current launch announcement dated November 18, 2026. It includes a model for completed recordings and two streaming models: Realtime for live transcripts and Realtime-Agent for software that replies to the speaker. All three support 32 languages.

The company says Realtime finalizes speaker labels when a turn ends. Realtime-Agent focuses on interactive turn-taking and offers Voice Focus for suppressing background speech. Those differences matter when choosing between a meeting caption panel and an automated phone conversation.

A transcription model supplies the listening layer. Teams building a complete voice assistant still need to decide how the application reasons, responds and handles interruptions.

Public record: AssemblyAI — STT 3.7 launch announcement · AssemblyAI — Model selection and feature comparison

Compare the billing unit as well as the price

These are the published base pay-as-you-go rates checked November 19. Recorded transcription is charged by audio duration; streaming is charged for the time the WebSocket session stays open. A silent connection can therefore affect the bill.

AssemblyAI STT 3.7 base list prices
ModelTypical jobBase rate
STT 3.7Completed recordings$0.21 per audio hour
STT 3.7 RealtimeLive captions and notes$0.30 per session hour
STT 3.7 Realtime-AgentVoice agents that answer$0.45 per session hour

Speaker labels and other optional features can add charges. These are transcription rates, not the total cost of a full voice-agent stack.

Public record: AssemblyAI — Model selection and feature comparison · AssemblyAI — Current API pricing

What changes for recorded interviews and meetings

The recorded-audio research post reports an expansion from 18 to 32 languages, including Marathi and Urdu among the additions. It also reports a 36% reduction in diarization error rate relative to Universal-3.5 Pro at launch. Diarization means assigning speech to the right speaker; it is a different measure from getting the words right.

These results come from AssemblyAI’s evaluations. They do not establish the same gain on every accent, recording setup or language combination. Promptwire has not independently benchmarked the models.

For an editor, speaker attribution deserves its own review. A fluent transcript can still put an important statement under the wrong name. Include brief interjections and overlapping voices when checking whether an upgrade helps your recordings.

Public record: AssemblyAI — Recorded-audio research and evaluations

December 8 changes defaults, not retirement status

AssemblyAI’s lifecycle documentation says that from December 8, 2026, existing accounts that omit speech_models and currently default to Universal-3.5 Pro or Universal-3 Pro will use STT 3.7 instead. New accounts created on or after November 17 already receive that default. Accounts defaulting to Universal-2 are excluded from the automatic change.

This is a default-model change. The two older recorded-audio models are deprecated, with retirement dates still to be announced. Separately, the documentation lists February 1, 2027 as the tentative retirement date for Universal-3.5 Pro streaming and the Universal-Streaming English and Multilingual models. Universal-3.6 Pro streaming remains generally available.

Check the actual model used by your requests before deciding whether either date affects the product.

Public record: AssemblyAI — Speech model versions and deprecations

Two output changes to check before switching

The migration guide says STT 3.7 preserves fillers, repetitions and false starts by default. Set disfluencies explicitly to false if the application expects cleaned transcripts. This deserves attention in subtitle editing, transcript search and any workflow comparing output with previous runs.

Automatic language detection also needs an explicit fallback choice. With only assemblyai-stt-3-7 in speech_models and language_detection enabled, the request does not fall back to another model. Include universal-2 in the list if you need that fallback for languages outside STT 3.7’s coverage.

A useful migration sample would combine a clean interview, a noisy group discussion and a recording that switches languages. Keep a reviewed reference transcript for each. Compare word corrections, speaker corrections and editing time separately, then check the charge with the optional features you actually enable. The practical win is a transcript that needs less repair in the application that uses it.

Public record: AssemblyAI — STT 3.7 migration guide

COMMON QUESTIONS

Questions readers ask

Is December 8 the retirement date for Universal-3.5 Pro recordings?

No. December 8 is the scheduled default-model switch for affected requests that omit speech_models. The recorded model’s retirement date is still to be announced.

SOURCES & NOTES

Follow the evidence.

Primary sources checked November 19, 2026. The current launch and research pages are dated November 18; the model lifecycle documentation records deprecations beginning November 17. Prices and performance claims are attributed to AssemblyAI. Promptwire has not independently tested the models. Evaluation suggestions are editorial analysis. Cover: AI-assisted conceptual illustration, not a product screenshot.

  1. AssemblyAI — STT 3.7 launch announcement
  2. AssemblyAI — Model selection and feature comparison
  3. AssemblyAI — Recorded-audio research and evaluations
  4. AssemblyAI — Speech model versions and deprecations
  5. AssemblyAI — STT 3.7 migration guide
  6. AssemblyAI — Current API pricing

Public sources support the factual background. Analysis and illustrative exercises are identified in the text. How we assess AI · Editorial standards

Spot a factual error or an outdated detail?