AssemblyAI documents a capability change: “# Fix: Duplicated Filler Words on Universal-3.5 Pro Universal-3.5 Pro transcripts generated with disfluencies: true could contain duplicated filler words,…”
# Fix: Duplicated Filler Words on Universal-3.5 Pro
Universal-3.5 Pro transcripts generated with disfluencies: true could contain duplicated filler words, for example "Uh uh" where the audio contains a single "uh"…
Read more
Universal-3.5 Pro transcripts generated with disfluencies: true could contain duplicated filler words, for example "Uh uh" where the audio contains a single "uh". This update significantly reduces duplicated disfluencies on internal benchmarks and is specific to Universal-3.5 Pro; other models are not affected.
This fix applies to new transcription requests going forward. Transcripts created before this fix keep their original text and are not retroactively corrected; if you need a corrected transcript for audio processed before this fix, resubmit the original audio file for reprocessing.
See the Filler Words docs →
Show less
Fix Pre-recorded Speech-to-Text
Entry date: 2026-09-01
# Fix: Reader Role Permissions Enforced for API Tokens
Read from the provider's text; the quote is the provider's exact lines.
AssemblyAI documents a capability change: “# Spoken Form Output for Universal-3.5 Pro Async Transcription Async transcription with universal-3-5-pro can now return spoken-form transcripts… Read more…”
# Spoken Form Output for Universal-3.5 Pro Async Transcription
Async transcription with universal-3-5-pro can now return spoken-form transcripts…
Read more
Async transcription with universal-3-5-pro can now return spoken-form transcripts.
Setting format_text: false preserves the transcript in spoken form rather than formatted text: numbers, dates, and other entities are left unnormalized, so a speaker saying "one hundred" returns as "one hundred," not "100," matching exactly what was said. This is useful for verbatim workflows like medical dictation, where preserving the literal spoken form matters.
Spoken-form output is available now for all accounts.
Read from the provider's text; the quote is the provider's exact lines.
AssemblyAI documents a capability change: “# Breaking: Async Default Model Changes to Universal-3.5 Pro on August 7 Read more Update (August 1, 2026): This change has been rescheduled to September 2,…”
# Breaking: Async Default Model Changes to Universal-3.5 Pro on August 7
Read more
Update (August 1, 2026): This change has been rescheduled to September 2, 2026, and the migration behavior has changed — requests pinned to universal-3-pro will return an error instead of being auto-routed. See the updated announcement for current details. The original entry follows below.
Requests that omit speech_models (or pass a legacy speech_model name) route to the default model. The default is universal-2 today and switches to universal-3-5-pro on August 7, 2026. On the same date, requests pinned to universal-3-pro are auto-routed to universal-3-5-pro.
If you rely on the default, your requests switch automatically that day, so set speech_models explicitly to pin your current model.
Read from the provider's text; the quote is the provider's exact lines.
# Universal-3.5 Pro Async: flagship pre-recorded speech-to-text across 18 languages
Universal-3.5 Pro is now live for pre-recorded audio — our most accurate async model yet, built for the natural complexity of real-world conversations, at $0.21/hr…
Read more
Universal-3.5 Pro is now live for pre-recorded audio — our most accurate async model yet, built for the natural complexity of real-world conversations, at $0.21/hr.
It transcribes across 18 languages with native mid-sentence code-switching, jointly solves transcription and speaker attribution instead of bolting on a separate diarizer, and takes domain context as input to sharpen accuracy on the terms that matter to you.
Read from the provider's text; the quote is the provider's exact lines.
