PARAS MADAN
All writing

Latest in AI

AssemblyAI STT-3.7 Migration: Check Your Defaults Before December 8

AssemblyAI's December 8 default switch changes more than the model name. Check language fallback, filler words and the model your transcription actually uses.

If your AssemblyAI integration leaves model selection to the account default, check it before December 8. A request that keeps working can still produce different transcripts, which matters when another system consumes the text.

AssemblyAI's November 18 announcement makes STT-3.7 generally available for recorded audio at $0.21 per hour. It supports 32 languages. This guide focuses on migrating recorded-audio workflows, not the separately launched realtime models.

Who changes on December 8?

The official migration guide says accounts currently defaulting to universal-3-pro or universal-3-5-pro will switch to STT-3.7 from December 8, 2026 when requests omit speech_models. Accounts defaulting to universal-2 will not automatically switch. Explicit model selection remains in effect until that model's retirement.

December 8 is a default change, not the retirement date. The model-selection documentation says Universal-3 Pro and Universal-3.5 Pro are deprecated, with retirement timing still to be announced. New accounts created from November 17 already default to STT-3.7.

I would inspect the actual request payload rather than rely on an application's settings label. Check background workers, upload handlers and any integration wrapper separately. If production and a test account were created at different times, identical requests without a model field are not a reliable comparison.

Make language fallback explicit

The model-selection documentation describes an important distinction. With automatic language detection and only STT-3.7 selected, there is no fallback. Include Universal-2 explicitly if you need broader language coverage. An explicitly supplied unsupported language_code, however, can trigger Universal-2 fallback even without it in the list.

STT-3.7 also takes priority wherever it appears in the list. Putting it second is not a way to reserve it as a backup for supported languages. Read speech_model_used from the completed result to confirm what ran.

For a multilingual application, I would include one representative recording from outside STT-3.7's supported set in the migration test. Otherwise, an English-only test could pass while missing the routing behavior that matters to a smaller group of customers.

Preserve the transcript style intentionally

The migration guide also changes the default for disfluencies: including STT-3.7 anywhere in the model list makes them enabled for the request, even if Universal-2 ultimately transcribes it. Set disfluencies explicitly to avoid relying on that default. Use false if the intended output removes fillers and similar interruptions.

The following is an illustrative request-body fragment for automatic language detection, broader fallback and cleaned text. It is not a complete runnable request or a tested result:

{
  "speech_models": ["assemblyai-stt-3-7", "universal-2"],
  "language_detection": true,
  "disfluencies": false
}

A cleaner transcript is not always the right output. For an editing workflow, preserving what was spoken may matter more than readability. Decide which behavior the product needs, then make that choice visible in configuration and tests.

Test the consumer of the transcript

My proposed rollout would compare a small, representative collection of recordings using explicit old and new model selections. Keep the audio and other settings constant. Include overlapping speakers, brief acknowledgments, names, numbers and any language mixtures common in your workload.

Review more than the transcription preview. Run the resulting text through the downstream summarizer, search index or extraction step. A change that looks harmless to a reader could affect a brittle matching rule. Record the actual model used alongside each result so fallback cases remain distinguishable.

The announcement also says the improved diarization system is already live across async speech models. Pinning an older speech model is therefore not a guarantee that every component of the processing pipeline remains frozen.

Before the default switch, choose an explicit rollout decision: migrate the tested path, or temporarily retain an available older model while resolving specific failures. Assign ownership for the eventual retirement notice too. An explicit model name buys control over this transition, not an indefinite exemption from future changes.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn API changes into clear migration guidance developers can use.

Book a 30-minute call.