Transcribe long audio
in three calls.

01Quickstart

Create a key in the console, then submit a file and poll for the result. Both calls use the same key.

curl -X POST https://api.vibevoice-asr.com/v1/jobs \
  -H "Authorization: Bearer $VVASR_KEY" \
  -F file=@meeting.mp3

{ "job_id": "8f14e45f-…", "status": "processing", "eta_seconds": 42 }

curl https://api.vibevoice-asr.com/v1/jobs/8f14e45f-… \
  -H "Authorization: Bearer $VVASR_KEY"

{
  "job_id": "8f14e45f-…",
  "status": "succeeded",
  "eta_seconds": 1,
  "result": {
    "text": "Let's take the full recording. And keep every speaker attached.",
    "segments": [
      { "start": 0.0, "end": 3.4, "speaker": "1", "text": "Let's take the full recording." },
      { "start": 3.4, "end": 6.1, "speaker": "2", "text": "And keep every speaker attached." }
    ]
  }
}

Call the API from your server, not from a browser. There are no CORS headers, so a fetch from page JavaScript will be blocked — and it would expose your key to anyone who opens devtools.

02Authentication

Every request needs Authorization: Bearer <key>. Keys look like vvasr_user_… and are shown exactly once, when you create them. Revoking a key takes effect immediately — there is no cache to wait out.

You can hold up to 10 active keys. Jobs belong to your account, not to the key that created them, so any of your keys can read any of your jobs. Someone else's job returns 404, never 403.

03Submit a job

POST /v1/jobs

Send a multipart/form-data form with the audio in it — either the bytes or a link, never both:

# upload the file
curl -X POST https://api.vibevoice-asr.com/v1/jobs -H "Authorization: Bearer $VVASR_KEY" \
  -F file=@meeting.mp3

# or let us download it
curl -X POST https://api.vibevoice-asr.com/v1/jobs -H "Authorization: Bearer $VVASR_KEY" \
  -F file_url=https://cdn.example.com/meeting.mp3

Returns 201:

{ "job_id": "8f14e45f-…", "status": "processing", "eta_seconds": 42 }

eta_seconds is an estimate of time to completion — use it to pace your first poll. The response also carries an Inference-Id header equal to the job_id.

04Fetch the result

GET /v1/jobs/{job_id}

Each segment carries start and end in seconds, the text, and speaker when the model identified one. Treat speaker as optional — it is a label, not a number, and it is omitted rather than guessed when the model gives nothing.

How often to poll

First poll after min(eta_seconds × 0.8, 2s), then back off exponentially up to 5s. A job that has not finished within 25 minutes is failed with deadline_exceeded, so that is your upper bound on waiting.

05Completion callbacks

POST {your callback_url} — sent by us, once per job

Leave an https URL in the callback_url form field when you submit, and we POST to it when the job reaches a terminal state, instead of making you poll for the news:

{ "event": "job.completed", "job_id": "8f14e45f-…", "status": "succeeded" }

Jobs that fail synchronously at submit time (4xx/5xx on the POST itself) don't trigger a callback — you already have the error in hand. Jobs we fail on your behalf (deadline_exceeded, dispatch_lost) do: for a client that isn't polling, that callback is the only signal it gets.

06Cancel a job

DELETE /v1/jobs/{job_id}

Stops a job that is still queued or processing and refunds the full reservation. Returns { "job_id": "…", "status": "cancelled" }. A job that already reached a terminal state returns 404 — there is nothing left to cancel.

07What you get charged

$0.0035 per audio minute, computed from the duration the transcription node actually decoded — not your file size, not how long the job took to run. Failed and cancelled jobs cost nothing.

Why your balance dips and then partly comes back

When you submit, we reserve an estimate based on the duration read from the file header, so a long job can't start against an empty balance. When it finishes we settle against the real decoded duration and return the difference. In the console you'll see this as a hold followed by a settle — the second one is usually a partial refund.

08Errors

Every error has the same shape: { "error": { "code": "…", "message": "…" } }. Some carry extra fields, noted below.

Failures after the job was accepted

These arrive as error.code inside a 200 poll response, with status: "failed": backend_failed, backend_timeout, bad_output, deadline_exceeded, dispatch_lost. None of them are billed — and each also fires your callback, if you left one.

09Limits

Maximum time to completion is a ceiling, not a typical duration — the estimate for your own job is the eta_seconds returned when you submit it. Work we predict cannot finish inside that ceiling is declined up front with 429 rather than accepted and timed out.

Transcription parameters (language hints, timestamp granularity) aren't available yet. If you need one, tell us — it helps us prioritise.