Bessel Consensus3 r1

Model and access

Transcribe audio

Use the dedicated API base URL below as BESSEL_API_URL. Create an ID once for a new recording and keep it for polling and recovery. Do not create a new ID to retry an ambiguous or failed request.

export BESSEL_API_URL="https://api.besselaudio.com"
export BESSEL_API_KEY="<private-consensus3-api-key>"
REQUEST_ID=$(python -c 'import uuid; print(uuid.uuid4().hex)')

curl "$BESSEL_API_URL/v1/audio/transcriptions" \
  -H "Authorization: Bearer $BESSEL_API_KEY" \
  -H "Idempotency-Key: $REQUEST_ID" \
  -F "model=Sonaflow/bessel-consensus3-r1" \
  -F "language=en" \
  -F "file=@audio.wav"

The response contains model, text, language, duration, and opaque profile_sha256 and runtime_sha256 version checksums. The returned text is normalized English: original punctuation and capitalization are not preserved. Empty output is possible. This version does not restore punctuation.

Asynchronous requests

Use the asynchronous job interface when possible. The service is configured to remain running during idle periods. Processing and queueing still take time, and deployments or instance recovery can require startup. A client timeout does not mean the job has stopped.

All API routes require authentication. Jobs are queued, running, then complete, failed, or interrupted. Completed jobs include a result object. If a response is lost, query the original job ID. Do not create a new job to retry a terminal or interrupted recording.

Only file, model, and language are accepted. Reference text, dataset names, prompts, and vocabulary hints are unsupported. Use one fixed English configuration for every evaluation dataset. Start with one worker.

Errors use HTTP 401 for authentication, 422 for invalid input, 413 for oversized uploads, 409 for an idempotency conflict, and 429 for a full queue. Keep keys out of public logs and source control.

Public evaluation evidence

The following are contributor-reported offline replay results, produced on October 22, 2026 (UTC) from archived public evaluation artifacts covering 76,544 inputs across ten test splits. Scoring used the unmodified official Open ASR Leaderboard English normalizer and scorer at commit 25e0b9e143b07eeaf3b50366c2525ddf0c0c9a6b. Local replay verified that this sealed release reproduces all 76,544 archived outputs.

The equal-weight mean of the ten rounded dataset scores is 4.152% (4.153480531994278% before per-dataset rounding). The official reference filter excludes 101 empty/ignored references, leaving 76,443 manifest rows; all 26 empty output strings in those rows are retained. Earnings22 uses official six-session regrouping.

Public benchmark data and results were used during development and reviewed before release selection. These scores are development evidence, not a held-out assessment. Pretraining overlap is not established. No fresh full-suite audio inference through this release's public endpoint has been performed, and RTFx is not reported. Offline replay does not establish fresh end-to-end API accuracy or speed. Maintainer private evaluation is pending; no independent verification or leaderboard rank is claimed.

See the API integration and evaluation submission for the adapter and aggregate evidence.

Evaluation data handling

The dedicated evaluation endpoint processes audio in memory and does not persist uploaded audio. Request IDs, audio digests, job status, and textual results are retained to support durable recovery; no fixed retention period is specified. This is not a claim of zero retention for derived text or metadata.

Sonaflow Inc. commits not to train on, log, store, retain, or inspect Open ASR Leaderboard evaluation audio throughout the processing path, including applicable contracted processing services. Private evaluation transcripts and derived labels will not be used for model development or tuning. Use the dedicated evaluation endpoint for leaderboard recordings. Evaluation results and leaderboard inclusion are subject to independent verification by the leaderboard maintainers.