Agent Skills: Whisper Lolo Transcription Jobs

Whisper Lolo Transcription Jobs

Overview

Run transcription asynchronously with Inngest, storing results in Postgres, and keep a provider abstraction for Whisper now and Voxtral later.

Job workflow

1. Emit recording.uploaded after successful Blob upload.
2. In Inngest, start transcribeRecording on that event.
3. Download audio from blob_url.
4. Chunk before transcription to fit provider limits.
5. Call the provider (Whisper now) and store text + segments.
6. Update statuses and handle errors with retries.

Provider abstraction

• Maintain a TranscriptionProvider interface with transcribe(audioUrl).
• Implement WhisperProvider now; leave VoxtralProvider stubbed.
• Avoid provider-specific logic in job orchestration.

Status discipline

• recordings.status: uploaded -> transcribing -> done/error.
• transcriptions.status: pending -> done/error.
• Persist timestamps for observability where possible.

Error handling

• Log failures with enough context to retry safely.
• Avoid duplicate transcription on retry (idempotency checks).

References to consult

• documentation/inngest-background-jobs.md
• documentation/inngest-demo-nextjs-full-stack.md
• documentation/openai-node-audio-example.md
• documentation/openai-speech-to-text.md
• documentation/openai-audio-api-reference.md