Long recordings

The asynchronous flow has four steps. The SDKs wrap them in one call (transcribe_file / transcribeFile).

#1. Create an upload

Tell the API how large the file is. Including its SHA-256 lets the API prove the stored bytes are exactly yours.

curl "$TWENTYSIX_BASE_URL/v1/uploads" \
  -H "Authorization: Bearer $TWENTYSIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"bytes\": $(stat -f%z meeting.m4a), \"sha256\": \"$(shasum -a 256 meeting.m4a | cut -d' ' -f1)\"}"

The response holds a short-lived, signed upload instruction:

{
  "id": "upl_…",
  "object": "upload",
  "state": "allocated",
  "bytes": 48213904,
  "upload": { "method": "PUT", "url": "https://…", "headers": { "x-upsert": "false" } },
  "expires_in": 3600
}

#2. Send the bytes

Send the file with the given method, URL and headers. The audio goes directly to storage in your project's region, without passing through the API servers.

curl -X PUT "$UPLOAD_URL" -H "x-upsert: false" --data-binary @meeting.m4a

Files above 50 MB can use the resumable_upload instructions in the same response instead (the tus protocol, 6 MB chunks). An interrupted resumable upload continues from where it stopped. The SDKs switch to it automatically.

#3. Complete the upload

curl -X POST "$TWENTYSIX_BASE_URL/v1/uploads/$UPLOAD_ID/complete" \
  -H "Authorization: Bearer $TWENTYSIX_API_KEY"

The API checks the size and checksum and measures the audio. The upload moves to verifying, then to complete with duration_seconds and channels, or to rejected with an error code such as audio_decode_failed. Poll GET /v1/uploads/{id} until it leaves verifying.

#4. Start the transcription

curl "$TWENTYSIX_BASE_URL/v1/transcriptions" \
  -H "Authorization: Bearer $TWENTYSIX_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: meeting-2026-10-20" \
  -d '{
    "audio": { "upload_id": "upl_…" },
    "language": "hi-IN",
    "mode": "codemix",
    "diarization": true,
    "webhook_endpoint_id": "wh_…"
  }'

The response is 202 Accepted with the transcription. Credit for the measured duration is reserved now and settled when it finishes.

{ "id": "tr_…", "object": "transcription", "state": "queued", "request": { "language": "hi-IN", "mode": "codemix", … }, "created_at": "…", "finished_at": null }

#Request fields

Unknown fields are rejected with unknown_parameter; nothing is silently ignored.

#5. Wait for the result

A transcription moves through queued → preparing → transcribing → finalizing and ends in one of:

Either register a webhook or poll GET /v1/transcriptions/{id} every few seconds. When the state is completed or partial, fetch the transcript:

curl "$TWENTYSIX_BASE_URL/v1/transcriptions/$ID/result" -H "Authorization: Bearer $TWENTYSIX_API_KEY"

#Cancel, delete, list

• POST /v1/transcriptions/{id}/cancel stops a transcription at the next safe point. Cancelled work is never charged.
• DELETE /v1/transcriptions/{id} deletes the audio and transcript of a finished transcription now, instead of waiting for retention. The transcription record remains with content_deleted_at.
• GET /v1/transcriptions?limit=20&state=completed,partial lists the project's transcriptions, newest first. Pass the response's next as after for the next page.

#Supported files

WAV, FLAC, MP3, M4A/AAC, OGG/Opus and WebM audio, up to two hours and 2 GB. Video files are coming soon.