FlowModels

Transcribe audio files: models, sizes and background jobs

Published on 11. November 2026

With FlowModels, you can turn recordings into text, for example phone calls, interviews or voice notes. Short files come straight back to you, and we process long recordings of up to eight hours in the background. On request, some models return timestamps. FlowModels doesn't currently offer speaker labels.

The models

Model File Additional information
groq/whisper-large-v3-turbo yes timestamps
groq/whisper-large-v3 yes timestamps
openai/gpt-transcribe yes text only

Only use models that appear in your response from GET /v1/models. There, transcription lists the supported modes and languages, and pricing.variants the prices per minute.

Try it in the Playground

You can also transcribe a recording of up to 24 MB and ten minutes in the dashboard without an API key: in the Audio tab of the Playground. Billing runs through your team's FlowModels credit. The step-by-step guide is in Try models in the Playground.

Transcribe a file

Send the file as multipart/form-data to POST /v1/audio/transcriptions:

curl https://api.flowmodels.one/v1/audio/transcriptions \
  -H "Authorization: Bearer at_your_key" \
  -H "Accept: application/json" \
  -F "model=groq/whisper-large-v3-turbo" \
  -F "language=en" \
  -F "file=@voice-note.mp3"

Your HTTP client sets the Content-Type with the boundary itself. You can't use a file previously uploaded through /files here via file_id.

Without further options, the transcription runs synchronously. That works for files up to 24 MiB and ten minutes. The response comes with HTTP 200 and contains text, usage and expires_at.

Long recordings in the background

For larger files, also set -F "background=true". Files up to 200 MiB and eight hours are then possible, though some models have their own, smaller limits. The API responds immediately with HTTP 202 and a job ID.

  1. Poll the job with GET /v1/audio/transcriptions/{id}.
  2. status is first queued, then processing. progress shows the progress between 0 and 1.
  3. As soon as status has the value completed, cancelled or failed, the job is finished.

POST …/{id}/cancel cancels a running job. A segment already running at the provider is still finished and billed. You can delete completed jobs with DELETE …/{id}.

Timestamps

Timestamps are off by default. Turn them on with timestamps=true, provided the model supports it. The result then contains segments with the start and end of each segment.

FlowModels doesn't currently offer speaker labels. No model supports diarize=true, and the API rejects this option.

If an option or language code doesn't fit the model, the API rejects the request with HTTP 422 before any costs arise at the provider.

What a transcription costs

Transcriptions are billed by the minute, and the price depends on the model and additional options. Before the first segment, we reserve the estimated total cost. Silence can cost money too. Groq charges at least ten seconds per call and per file segment. What was actually billed is shown in usage.cost.

Retention

You can fetch finished results for 24 hours after completion, and expires_at states the exact time. We delete your uploaded audio file after completion, cancellation or failure.

Live transcription

For ongoing conversations, there are also live sessions over WebSocket once live models appear in your model list, see Use live transcription.

If something goes wrong

  • 503 audio_disabled: Transcription isn't available right now.
  • 503 audio_model_unavailable or 503 audio_pricing_unavailable: There is currently no access to or no price for the model. Use a different model or try again later.
  • 422: The file is too large or too long for the synchronous route, or an option doesn't fit the model. Set background=true or check the options.
  • Job failed, but text is present: After a failure or cancellation, already confirmed partial results may be in the job.

Was this article helpful?

Back to Models, limits & errors