Skip to content

Audio captioning models

Symphony Notes, what it hears, and how it differs from the transcription models.

One model captions audio: Symphony Notes, on POST /v1/chat/completions. Send the id hyphenated as symphony-notes. The model field is case-insensitive, but the spaced name does not resolve.

Symphony Notes

A general-purpose audio captioner. It hears speech, speakers, and background sound together and returns a plain-language description, not a transcript.

FieldValue
Model valuesymphony-notes
ModalitiesAudio in, text out
EndpointPOST /v1/chat/completions
InputOne clip as an input_audio content part

What it hears

  • Speech content. What was said, described rather than transcribed word for word.
  • Speakers and emotion. Who is talking, and the feeling in how they say it.
  • Delivery. Pace, volume, and the way a line is carried.
  • Acoustic environment. The room, the street, the hall behind the voice.
  • Non-speech sound. Music, a buzzer, footsteps, a door.

Caption content varies with the clip, so don't parse it for expected fields. To get stable fields, request structured output.

Clip limits

Two limits cap a clip near four minutes.

LimitValueAudio it allows
Context4096 tokens, audio at about 13 tokens per second~5m15s, before the caption takes its share
Media size8 MiB, measured after transcoding~4m22s

Over the media limit the request returns 413; over the context window it cannot run at all. Neither limit degrades gracefully, so segment long recordings and caption each piece. Shorter clips generally get more detailed descriptions, and per-segment captions index against timecodes.

How it compares

Several models take audio in. Pick by what you want back.

ModelReturnsUse it when
Symphony NotesProse describing the recordingThe sounds matter as much as the words
RhythmA transcript, with timings and speakersYou need the words themselves
SymphonyA spoken or written replyYou want the model to answer

Symphony Notes and Rhythm both take audio and return text. Use Symphony Notes for a description of the recording: asked for a transcript, it returns prose about the speaking. Use Rhythm for the words themselves: it reports nothing about the music.

See the model catalog for the live list and pricing for current rates.