Media Transcript Extractor

Media Transcript Extractor

Workflow

1. Confirm the input is a local audio/video path. If the user supplied a relative path, resolve it against the user's original task workspace.
2. Preserve that workspace before launching Python. The local Python runtime can fail when it initializes inside a restricted workspace, so pass an absolute output path and let the script switch its runtime directory internally. Never infer the output folder from /tmp.
3. Prefer the bundled launcher. It preserves the caller's workspace for relative input/output paths while starting Python from a neutral runtime directory:

/Users/otonashic./.codex/skills/media-transcript-extractor/scripts/transcribe_media \
  "/path/to/media.mp3" \
  --out-dir "/absolute/task/workspace/outputs/media-transcripts"

4. If the language is known, pass it explicitly:

/Users/otonashic./.codex/skills/media-transcript-extractor/scripts/transcribe_media \
  "/path/to/media.mp4" \
  --language fr \
  --out-dir "/absolute/task/workspace/outputs/media-transcripts"

5. For sourcing/product/geography videos, add likely terms as a prompt to improve brand and place names:

--prompt "Terms: Fujian, Putian, Jinjiang, Guangdong, Guangzhou, Dongguan, Zhejiang, Wenzhou, Anta, Xtep, 361, Peak."

6. Validate coverage before reporting success:
   • Read the script JSON summary or final console output.
   • Check that last_end is near the media duration.
   • If the transcript seems incomplete, add a better --prompt or review the audio around the missing section.
   • Do not switch to turbo unless the user explicitly requests that model.

Defaults

• Default model: medium, regardless of media duration.
• Never auto-select turbo based on duration. Use turbo only when the user explicitly requests it.
• Warn that long recordings can take substantial time with medium. If the user does not want to use this skill for long media, stop instead of silently changing models.
• Default execution: local faster-whisper; do not use Feishu/Lark Minutes unless the user explicitly requests it.
• Default outputs:
  • Markdown report: *.md
  • Plain transcript: *.txt
  • Timestamped segments: *.segments.txt
  • Machine-readable details: *.json

Local Environment

The script auto-detects the existing faster-whisper runtime at:

/Users/otonashic./.claude/skills/tiktok-analyzer/.venv/bin/python

It also prefers cached Hugging Face models on this machine:

~/.cache/huggingface/hub/models--mobiuslabsgmbh--faster-whisper-large-v3-turbo
~/.cache/huggingface/hub/models--Systran--faster-whisper-medium

Use environment variables only when the defaults stop matching the machine:

FASTER_WHISPER_PYTHON=/path/to/python
FASTER_WHISPER_MODEL=/path/to/model

Output Guidance

• Treat raw ASR as evidence, not polished copy.
• Do not silently invent uncertain brand/place names. If cleaning the transcript, keep a note for corrected or uncertain terms.
• For short-form videos, preserve timestamps; they make it easier to compare suspicious words against the audio.