Speech to text over REST

POST /v1/audio/transcriptions transcribes a finished audio file in one request. It follows the OpenAI transcriptions shape, so point an existing OpenAI client at it by changing the base URL and the model. Models lists the models that transcribe; the examples below use rhythm.

Transcribe a file

Upload the audio as multipart/form-data.

curl https://api.sprag.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPRAG_API_KEY" \
  -F [email protected] \
  -F model=rhythm

{
  "text": "Thanks, everyone, for joining. Let us start with the roadmap.",
  "usage": { "input": { "audio": { "duration": "5" } } }
}

usage is abbreviated in the examples on this page.

Request fields

Response fields

json and verbose_json return a JSON object with the fields below. text returns the transcript as a plain text body, and srt and vtt return a rendered caption file.

Stream a transcript

To receive the transcript while the file is still being transcribed, set stream=true. The response is a stream of server-sent events.

curl -N https://api.sprag.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $SPRAG_API_KEY" \
  -F [email protected] \
  -F model=rhythm \
  -F stream=true

data: {"type":"transcript.text.delta","delta":"Thanks,"}

data: {"type":"transcript.text.delta","delta":" everyone, for joining."}

data: {"type":"transcript.text.delta","delta":" Let us start with the roadmap."}

data: {"type":"transcript.text.done","text":"Thanks, everyone, for joining. Let us start with the roadmap.","usage":{"input":{"audio":{"duration":"5"}}}}

Each transcript.text.delta event carries the next piece of the transcript in delta. The stream ends with one transcript.text.done event. The transcript.text.done event carries the full transcript in text and the same usage object a non-streamed response returns.

stream applies with response_format set to json or text. If you send stream=true with verbose_json, srt, or vtt, the request returns HTTP 400. If transcription fails after the stream starts, the last event's data is an error object, and no transcript.text.done event follows.

Timestamps and speaker labels

Set response_format=verbose_json when you need more than the transcript string. That response adds task, duration, a segments array, and a words array, with start and end times in seconds on every entry.

When captions are the end product, request srt or vtt instead and skip the client-side assembly; the response body is the rendered cue file.

Speaker labels attach to the word objects. Add speaker_labels=true and each attributed word gains a speaker index; num_speakers sets the count when you know it. Speaker labels need verbose_json and, if you send a granularity list, word in it.

Timestamps and diarization cover the payloads.

Errors

Full schemas are in the transcriptions reference.