Changelog.

Multi-backend model routing

Jobs can specify model_preference ("whisper", "parakeet", or "auto") to route to the best available transcription backend. Parakeet TDT 0.6B v2 is English-only and faster than Whisper on compatible hardware.

Batch job submission

POST /api/v1/jobs/batch accepts multiple transcription jobs in one request, returning per-item results including idempotency-key deduplication.

Language routing

Jobs can declare an ISO 639-1 language code to route non-English audio to the correct backend. When omitted, the coordinator defaults to "en".

Model catalog endpoint

GET /api/v1/models lists supported backends, model identifiers, language support, and quality tiers so clients can discover options before submitting work.

Async job ingestion

Accepted jobs return 202 while audio ingestion runs asynchronously. Ingestion errors surface on the job status endpoint rather than at submit time.

Shard progress in one call

GET /api/v1/jobs/{id}?include=shards returns shard-level progress without a separate request to /shards.

Client endpoints

• POST /api/v1/jobs — submit a transcription job
• GET /api/v1/jobs — list jobs with pagination and status filter
• GET /api/v1/jobs/{id} — get job status
• GET /api/v1/jobs/{id}/result — get completed transcript
• GET /api/v1/jobs/{id}/shards — inspect shard-level breakdown
• GET /api/v1/jobs/{id}/events — full audit trail for a job
• GET /api/v1/models — list available backends and models

Auth model

Client requests use Authorization: Bearer <api_key>. API keys are SHA-256 hashed at rest. Keys are managed through partner records and never stored in plaintext on the coordinator.

Stability policy

Breaking changes increment the major version. Additive fields in JSON responses are backward compatible. The full engineering changelog is maintained in the repository; this page lists client-visible entries only.