Quotaflow
llms.txtOpenAPIDashboard
OpenAI-compatible

Audio Transcriptions

When to use this page

Use this page to turn a recorded WAV file into text with the OpenAI-compatible file transcription request shape.

Endpoint

POST https://api.quotaflow.ai/openai/v1/audio/transcriptions

The request body is multipart/form-data.

Supported models

Check the exact model ID with GET /openai/v1/models using the same key. Muse Voice is a separate audio product, not a Muse text model; catalog membership does not guarantee live capacity.

Request fields

FieldRequiredValues
fileyesOne WAV file: RIFF/WAVE, mono, 16-bit integer PCM, 16 kHz or 24 kHz, at most 10 minutes, request body at most 32 MB
modelyesA supported model above
response_formatnojson (default) or text

Other OpenAI transcription fields (language, prompt, temperature, stream, timestamp_granularities[], include[], chunking_strategy) are refused with 400: this model returns the plain transcript, with no segments, word timestamps, speaker labels or streaming. Convert other audio formats to WAV first, for example ffmpeg -i input.mp3 -ac 1 -ar 16000 -sample_fmt s16 output.wav.

Curl

curl https://api.quotaflow.ai/openai/v1/audio/transcriptions \
  -H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
  -F model=meta/muse-voice-transcribe-1.0 \
  -F file=@recording.wav

Response shape

With response_format json:

{
  "text": "The quick brown fox jumps over the lazy dog.",
  "usage": { "type": "duration", "seconds": 2 }
}

usage has the shape OpenAI uses for models billed by audio duration. usage.seconds is the length of your audio measured from the WAV file, rounded down to whole seconds: the seconds this request is billed for (see Billing). A recording shorter than one second shows 0. Treat usage as optional: an answer recorded before usage was added that is returned again for a repeated request does not carry it.

With response_format text the body is the transcript as text/plain, with no usage.

Billing

You are billed for the length of the audio you upload, measured from the WAV file itself and rounded down to whole seconds, at the model's published per-hour audio price. A recording shorter than one second is billed zero seconds. A request that fails before a transcript is returned is not billed. A file whose length cannot be measured (not a WAV in the format above, or a WAV whose audio data chunk is not its last section, or that has an incomplete or odd-sized chunk; RIFF padding is read differently by different readers, so a file with an odd-sized metadata chunk is refused and must be re-exported) is refused with 400 before it is sent for transcription.