Skip to main content
Add up to 10 clips, each at most 30 minutes long, under assets.audios to decide on what something sounds like: the intent of a voicemail, whether a call followed a script, or how clearly a voiceover is delivered. The model hears the audio itself. You do not need a transcript. Use state to say what each clip is.

Route a voicemail

Illustrative response for a 40-second voicemail:
Response

Check a call against a script

Use a JSON state to give the script with the recording. This example also scores delivery.

Audio inputs

For long recordings, upload with the direct upload flow and pass the asset_id. A 30-minute WAV file is larger than the URL and base64 caps. MP3 at 128 kbps is about 29 MB for 30 minutes.

How much audio fits

If a request does not fit the context window, the API returns 400 with context_length_exceeded. See Context window.

Audio with video

The model hears the soundtrack of each video in assets.videos. Send a video’s sound as a separate clip only when you have it apart from the video. The response has the same data.answers and data.usage fields as a text request.