Speech-to-text - AiQL Handbook

The speech-to-text model turns spoken audio into transcript text. Convert on the Middleware API uses it for audio files and for the soundtrack of video. The transcript then follows the same split, shape, and extract path as any other source.

Parameters

The production ASR model is language-model speech recognition. It is multilingual. Language can be pinned per ingest.

Parameter	Value
Kind	Language-model automatic speech recognition
Input	Audio (`mp3`, `wav`, `m4a`) and video soundtrack (`mp4`, `mov`)
Languages	Multilingual. ISO language codes can be pinned. Each pinned code produces its own transcript page
Biasing	Domain terms and proper nouns can be supplied so recognition prefers them
Output	Transcript text into convert

Configuration

Setting	How it is configured
Convert	Audio and video sources are transcribed before split, shape, and extract.
Language	Callers can pin one or more ISO codes. Unpinned language is auto-detected where the model allows it.
Serving	AiQL selects the checkpoint. Customers do not pick a speech-to-text model.

⌘I