AssemblyAI: Universal-3.5 Pro

Universal-3.5 Pro is AssemblyAI’s speech-recognition model for converting recorded speech into text. In its Sync profile, a short audio clip produces a completed transcript in one response. This suits voice messages, push-to-talk input and discrete conversational turns where the application already knows when the recording should end.

Prepare the audio as a bounded clip. AssemblyAI’s Sync requirements accept durations from 80 milliseconds to 120 seconds, with a maximum file size of 40 MB. The supported formats are WAV or raw signed 16-bit little-endian PCM. Audio can be mono or stereo, with stereo down-mixed internally. Supported sample rates are 8, 16, 22.05, 24, 32, 44.1 and 48 kHz. A WAV header supplies the sample rate and channel count; raw PCM needs those values supplied separately in the transcription configuration.

Capture the complete utterance and keep the beginning and ending of the speech intact. A useful original task is a short product-support voice message containing a product name, an order identifier and a request. Retain the source clip so the transcript can be checked against those important details. For a longer recording, prepare a workflow appropriate to long-form transcription rather than treating the short-clip duration limit as a model-wide audio capacity claim.

The result contains the full transcript, a word list with per-word confidence and an overall confidence between zero and one. It also provides the submitted audio duration and a request identifier, making it possible to keep the text associated with its original clip. Use word confidence to locate passages that deserve listening review, especially names, numbers and domain vocabulary that matter to the application.

Word timestamps are optional and off by default. When requested, each aligned word can carry its start and end in milliseconds relative to the clip’s beginning. The maker’s profile leaves those values absent when a word cannot be aligned, rather than inventing an approximate timing. Timestamp accuracy is best for English audio. Retain missing timings explicitly when preparing word highlighting, a searchable waveform or subtitle alignment, and compare the text with the recording before using it as a finished caption track.

A contextual prompt describes what the recording is about. It can name a domain, a scenario or specific details already known to the application, and its maximum length is 6000 characters. Start with a short description that fits the audio, such as a technical-support conversation about a particular device. The prompt supplies recognition context; it is not a request to rewrite the speaker’s message into a different document.

A separate keyterm list biases recognition towards specific vocabulary. The Sync profile accepts up to 100 terms and 8000 characters across that list. Use exact spellings for unusual names, acronyms or product terms, and leave ordinary well-recognised words out of the list. Large or unnecessary term lists can cause overcorrections, so focus the list on words that matter in the recording. The context description and vocabulary list can be used together.

Prior conversation can also help interpret a short reply. Supply previous utterances in chronological order, from oldest to newest, while the current audio remains the turn being transcribed. The documented context allowance is 500 turns or 16,000 characters in total. If it exceeds those caps or the available model context, the oldest entries are removed first. This is useful when a brief answer contains an identifier or a reference whose meaning comes from the preceding question.

Language selection steers recognition towards one or more expected languages. The current Sync interface accepts 32 language codes and defaults to English when no selection is supplied. For code-switching audio, list the expected languages. A custom contextual prompt replaces the managed default and overrides the separate language-code steering, so name the expected languages inside that prompt when both language and scenario context are needed. Keep the transcript in the languages actually spoken rather than treating transcription as a translation instruction.

The model also appears in other AssemblyAI transcription workflows, while the Sync profile’s short-clip requirements and result fields belong to this particular interface. The maker’s Dictation workflow adds a separate text-cleanup stage; a transcript and a rewritten message are different artifacts. Preserve the recognised speech if a later editing or summarisation step produces a cleaned version, so the original wording remains available for comparison.

Useful deliverables include a transcribed voice message, a completed turn for an assistant, short searchable audio text or a timed word list for a review interface. Check the decisive words against the recording, preserve any missing alignments and keep the clip linked to the returned text. The recognition result gives the next application step a source-backed transcription; answering the request or editing the message belongs to the surrounding workflow.