Agent Skills: Whisper Lolo Audio Ingest

Whisper Lolo Audio Ingest

Overview

Implement long-form browser recording with chunked storage in IndexedDB and direct uploads to Vercel Blob, without serverless upload limits.

Recording workflow

1. Initialize MediaRecorder with a supported mime type.
2. Start with MediaRecorder.start(timeslice) to emit chunks.
3. On dataavailable, persist each chunk to IndexedDB.
4. On stop, rehydrate chunks and assemble a final Blob.
5. Clear stored chunks after a successful upload.

Storage guidance

• Do not keep full audio in RAM; always store chunks in IndexedDB.
• Use idb-keyval for simple storage of Blob chunks.
• Guard against empty chunks; some browsers emit zero-size data.

Upload workflow (client uploads)

1. Use upload() from @vercel/blob/client.
2. Generate tokens via a server route using handleUpload.
3. Persist blob_url and update status to uploaded after completion.
4. Use onUploadProgress for UX feedback on large files.

Non-negotiable constraints

• Never upload audio via a Next.js API route.
• Do not wait for transcription inside HTTP requests.
• Chunk before transcription; upload only after assembly.

Common pitfalls

• Check MediaRecorder.isTypeSupported() before selecting mime type.
• Resume/pause should not break chunk order in IndexedDB.
• Ensure onUploadCompleted works locally only with a tunnel or VERCEL_BLOB_CALLBACK_URL.

References to consult

• documentation/mediarecorder-mdn.md
• documentation/web-dictaphone-mdn.md
• documentation/idb-keyval.md
• documentation/mediarecorder-examples-mozdevs.md
• documentation/vercel-blob-client-uploads.md
• documentation/vercel-blob-sdk.md
• documentation/vercel-blob-examples.md