Skip to main content
Follow this guide to move from universal-3-pro, universal-3-5-pro, or universal-2 to assemblyai-stt-3-7.

Update your speech_models

If you don’t set speech_models Your default model may change. Check which model your requests use on the usage page, or send a test request and check which model processed it, as described in Identify the model used. Then:
  • universal-3-5-pro or universal-3-pro: your requests will automatically use assemblyai-stt-3-7 starting December 8, 2026. This is a change to the default model, not the retirement of universal-3-5-pro or universal-3-pro. See Behavioral changes.
  • universal-2: you won’t be upgraded automatically. To upgrade, set speech_models to ["assemblyai-stt-3-7", "universal-2"]. If all your audio is in languages AssemblyAI-STT-3.7 supports, ["assemblyai-stt-3-7"] is enough.
If you set speech_models You won’t be upgraded to the latest model until that model is retired. To migrate now, change the model in speech_models to assemblyai-stt-3-7: Keep universal-2 in the list if you need languages AssemblyAI-STT-3.7 doesn’t support. If all your audio is in languages AssemblyAI-STT-3.7 supports, ["assemblyai-stt-3-7"] is enough.

Behavioral changes

Model fallback

Previously, AssemblyAI fell back to another model by default. For example, if you used automatic language detection and set speech_models to ["universal-3-pro"] or ["universal-3-5-pro"], and AssemblyAI detected a language that the model doesn’t support, it automatically transcribed with universal-2 instead. Now, when you set speech_models to ["assemblyai-stt-3-7"], whether AssemblyAI falls back to another model depends on how you set the language:
  • language_detection: true: AssemblyAI always transcribes with AssemblyAI-STT-3.7 and never falls back to another model. To fall back to Universal-2 for detected languages that AssemblyAI-STT-3.7 doesn’t support, set speech_models to ["assemblyai-stt-3-7", "universal-2"].
  • language_code: If you set a language code that AssemblyAI-STT-3.7 doesn’t support, AssemblyAI falls back to Universal-2.
To keep the previous fallback behavior, explicitly set speech_models to ["assemblyai-stt-3-7", "universal-2"]:
AssemblyAI-STT-3.7 takes priority wherever it appears in speech_models. For example, if you add it after an existing model, as in ["universal-2", "assemblyai-stt-3-7"], AssemblyAI-STT-3.7 transcribes every language it supports, and Universal-2 handles the rest. If speech_models doesn’t include assemblyai-stt-3-7, the first model in the list, from left to right, that supports the language is used, as before. See Transcribe all 99 languages by combining speech models for details.

Disfluencies

The disfluencies default depends on the speech model. AssemblyAI-STT-3.7 keeps disfluencies by default. Universal-3.5 Pro, Universal-3 Pro, and Universal-2 remove them by default. To get the same behavior on every model, set disfluencies explicitly. See Verbatim.
On AssemblyAI-STT-3.7, disfluencies: true keeps more than filler words. It also preserves repetitions, stutters, false starts, and informal contractions such as “gonna” and “kinda”. On Universal-2, it keeps filler words only. If you process transcript text downstream, check your audio with the new output before you migrate.
When speech_models includes assemblyai-stt-3-7 in any position, disfluencies defaults to true for the whole request. This includes lists such as ["assemblyai-stt-3-7", "universal-2"] and ["universal-2", "assemblyai-stt-3-7"], even when Universal-2 transcribes the audio. To keep the default you had on universal-3-pro or universal-3-5-pro and remove disfluencies, set disfluencies to false:

Code switching

AssemblyAI-STT-3.7 handles code switching natively, so you don’t need to turn it on. The code_switching and code_switching_confidence_threshold options in language_detection_options apply only to Universal-2. If your requests set them, they have no effect when AssemblyAI-STT-3.7 transcribes the audio. Keep them only if Universal-2 is in your list and handles some of your languages. See Code switching.

Restrict the languages

To limit transcription to a known set of languages on AssemblyAI-STT-3.7, use language_codes instead of language_detection_options.expected_languages:
On AssemblyAI-STT-3.7, language_codes accepts any number of supported languages, and English isn’t required. A one-element list such as ["es"] behaves like setting language_code to "es". On Universal-2, language_codes accepts at most two languages, and one of them must be "en". language_codes can’t be combined with language_code or language_detection. See Constrain the language set.