feat(transcribe): add local Whisper transcription helper

[abman4444]

What

Adds helpers/transcribe_whisper.py — a drop-in alternative to the ElevenLabs Scribe transcribe.py helper for users who don't have an API key or prefer a free, fully offline option.

Why

The existing pipeline requires an ElevenLabs API key for transcription. Users on free-tier accounts or those who want a zero-cost setup have no fallback. Local Whisper fills that gap.

How it works

• Uses openai-whisper with word_timestamps=True to produce word-level timestamps
• Output JSON matches the Scribe envelope shape (words, text, language_code, alignment) so pack_transcripts.py, render.py, and the rest of the pipeline work unchanged
• Caches transcripts per source file — skips re-transcription on repeat runs (same behaviour as Scribe helper)
• Extracts mono 16kHz WAV via ffmpeg before passing to Whisper (same as the Scribe path)
• Supports all Whisper model sizes (tiny → large); defaults to medium (good balance of speed and accuracy for English)

Usage

# Basic
python helpers/transcribe_whisper.py my_video.mp4

# Custom model and language
python helpers/transcribe_whisper.py my_video.mp4 --model large --language en

# Custom output directory
python helpers/transcribe_whisper.py my_video.mp4 --edit-dir /path/to/edit

Reviewer notes

• No changes to existing files — purely additive
• Requires openai-whisper (pip install openai-whisper) and ffmpeg on PATH, both already listed as setup dependencies
• Word-level timestamp accuracy is slightly lower than Scribe (Whisper drifts ~50–100ms) but within the cut-padding working window defined in SKILL.md

---

Summary by cubic

Add a local Whisper transcription helper that outputs Scribe-compatible JSON, enabling offline, zero-API-key transcription without changing the rest of the pipeline. Adds caching and model selection for better control and speed.

• New Features

  • Added helpers/transcribe_whisper.py using openai-whisper with word_timestamps=True.
  • Emits Scribe-compatible JSON (words, text, language_code, alignment) so pack_transcripts.py and render.py work unchanged.
  • Extracts mono 16 kHz WAV via ffmpeg before transcription.
  • Caches transcripts per source file to skip repeat runs.
  • Supports all Whisper models (tiny → large); defaults to medium. Optional --language.

• Dependencies

  • Requires openai-whisper and ffmpeg available on PATH.

^{Written for commit ff6884c. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}