0.12.0 regression: local .ogg transcription via Groq returns 403, while 0.11.1 works

[mrpl327]

Summary

@steipete/summarize 0.12.0 fails to transcribe a local .ogg audio file with Groq in my environment, while 0.11.1 works on the same machine with the same GROQ_API_KEY and the same audio file.

This looks like a regression in the Groq transcription path introduced/exposed between 0.11.1 and 0.12.0.

Environment

• summarize 0.11.1 → works
• summarize 0.12.0 → fails
• OS: Linux
• Node: v22.22.1
• Input: local .ogg file (audio/ogg)
• Groq API key is valid

What works

1) summarize@0.11.1

summarize "/path/to/file.ogg" --extract-only --json

Example output shape:

Transcript:
<successful transcript text from the local audio file>

2) Direct curl to Groq transcription endpoint

curl https://api.groq.com/openai/v1/audio/transcriptions \
  -H "Authorization: Bearer $GROQ_API_KEY" \
  -F "file=@audio.ogg" \
  -F 'model=whisper-large-v3-turbo'

Example response shape:

{"text":"<successful transcript text>", ...}

What fails

summarize@0.12.0

summarize "/path/to/file.ogg" --extract-only --json

Error:

Failed to transcribe media (Groq transcription failed; falling back to local/AssemblyAI/Gemini/OpenAI: Groq transcription failed (403): {"error":{"message":"Forbidden"}}; yt-dlp transcription failed: Groq transcription failed: Groq transcription failed (403): {"error":{"message":"Forbidden"}})

Important A/B result

Same machine, same file, same key:

• 0.11.1 ✅ works
• 0.12.0 ❌ fails with 403 Forbidden

Extra debugging

I also reproduced the failure with a minimal Node fetch + FormData + Blob upload using the same Groq endpoint and same model:

const form = new FormData();
form.append('file', new Blob([bytes], { type: 'audio/ogg' }), 'audio.ogg');
form.append('model', 'whisper-large-v3-turbo');
await fetch('https://api.groq.com/openai/v1/audio/transcriptions', {
  method: 'POST',
  headers: { Authorization: `Bearer ${key}` },
  body: form,
});

This also returns:

403 Forbidden

while curl succeeds.

So this may be a multipart upload compatibility regression (or a path that now consistently hits a Groq-incompatible request shape) rather than a bad key.

Why I suspect 0.12.0 specifically

The 0.12.0 release notes mention:

• adding AssemblyAI as a first-class transcription provider
• refactor remote fallback ordering

That matches what I’m seeing: 0.12.0 appears to hit a Groq-first path that is broken in my environment, while 0.11.1 does not.

Request

Could you check whether the Groq transcription upload path changed between 0.11.1 and 0.12.0, especially around:

• multipart FormData construction
• Blob/filename/content-type handling for local .ogg
• provider/fallback ordering changes that make Groq the first path more consistently

I can provide more details if helpful.