Skip to content

Follow-ups from the whisper-1 removal (#48) #57

Open

Description

@rianvdm

When I investigate an audio rejection, the check script (npm run check:chunks -- '<url>' --transcribe, which the failure alert now tells me to run) can report FAILED for uploads that gpt-transcribe accepts. In a 2026-11-12 test with a stitched MP3 (30 seconds at 256 kbps followed by 20 minutes at 32 kbps, each piece with its own ID3 tag and Info frame), ffmpeg reported Header missing on part 1 and the script failed it. Both uploads returned HTTP 200 with transcripts.

Feed-path logs also record more than they should. The RSS fallback log records the full feed URL (apple-podcasts.ts) and the Podcast Index lookup-failure log records every response header (podcast-index.ts). If a feed URL contains a token, the log keeps it.

Outcome: the stitched-MP3 reproduction passes the check while real decode errors still fail it, and feed-path logs carry only hostnames and allowlisted response headers, the rule the audio path already follows.

Follow-up to #48 (merged in 1743b90).

Changes

  • Reproduce the interior-metadata error in the MP3 decode check (check-chunks.ts), isolate its cause, and let expected metadata through while malformed audio still fails. Decoding each piece separately is one option.
  • Restrict feed-path logging to hostnames and allowlisted response headers, starting with the two events above.

Smaller items

  • Add readVbrHeader cases for stereo and MPEG-2/2.5 side-info offsets, a header without the frame-count flag, and a truncated count (test/mp3.test.ts).
  • Add a regression test for the whole-file fallback when a small MP4's sample table points past the end of the file. The current fallback test uses a fragmented MP4.
  • When a rejected last chunk is skipped, partialReason repeats the rejection in parentheses ("was rejected by gpt-transcribe (gpt-transcribe rejected the audio)", transcription.ts). Put OpenAI's error message there, or drop the parenthesis.
  • The KV-fallback timeout test (job-timeout.test.ts) prints the DO-update error from updateJobStatusDO into the test output. Capture and assert it.
  • Document in CLAUDE.md how the check script fetches audio differently from the Worker: it follows redirects with its own HEAD request and needs range responses, while the Worker's default path goes through resolveAudioUrl() and validateAudioUrl() and can fetch a small file whole (the admin overrides skip both).
  • The README says monitoring runs every 8 hours; wrangler.toml runs it every 2.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions