agentleFS
pengasuziepengasuzie/claude-code-starter/.claude/skills/voice-to-text/SKILL.md

voice-to-text

Skill1 starsChanged 9 months agoMITView on GitHub

Transcribe audio/video files to text with speaker identification. Use when the user mentions "transcribe," "transcription," "voice to text," "audio to text," "meeting notes," "interview transcript," "diarize," "speaker identification," or provides an audio/video file path and wants it converted to text.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. About 0 are common, 5 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    whisperX: must be installed globally via pipx: pipx install "whisperx @ git+https://github.com/m-bain/whisperX.git" --python python3.13

    line 12
  2. Very specific:

    Hugging Face token: must be saved at ~/.cache/huggingface/token (for pyannote speaker diarization)

    line 13
  3. Very specific:

    If the user hasn't provided a file path, first check ~/Downloads/ for the most recently added .m4a file using:

    line 20
  4. Very specific:

    If a file is found, show the filename and ask the user to confirm it's the right one before proceeding.

    line 26
  5. Distinctive:

    Ask only if not obvious from context:

    line 32
  6. Very specific:

    Always use --compute_type int8 on Mac (float16 is not supported on CPU)

    line 54
  7. Very specific:

    Always set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 (PyTorch 2.6+ compatibility with pyannote)

    line 55
  8. Distinctive:

    Read the output file

    line 62
  9. Very specific:

    If speaker labels are generic (SPEAKER_00, SPEAKER_01), ask the user if they want to replace them with real names

    line 64
  10. Distinctive:

    If the user wants cleaner output, offer to:

    line 68
  11. Distinctive:

    After the transcript is ready, ask:

    line 74
  12. Very specific:

    If the user says yes, invoke the voice-briefing skill with the transcript file path.

    line 78
  13. Distinctive:

    If the user just wants plain transcription (no speaker identification), drop the diarize flags:

    line 92

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Voice to Text

You are a transcription assistant. Your job is to transcribe audio and video files into clean, readable text with speaker identification using whisperX.

Prerequisites

When Invoked

Step 1: Get the audio file path

If the user hasn't provided a file path, first check ~/Downloads/ for the most recently added .m4a file using:

ls -t ~/Downloads/*.m4a 2>/dev/null | head -1

If a file is found, show the filename and ask the user to confirm it's the right one before proceeding. If no .m4a is found in Downloads, or the user says it's not the right file, ask them for the path.

Supported formats: .m4a, .mp3, .wav, .flac, .ogg, .webm, .mp4, .mkv, .avi

Step 2: Determine parameters

Ask only if not obvious from context:

ParameterDefaultWhen to ask
Number of speakers2If unclear (e.g., meeting vs interview vs monologue)
LanguageenIf file might not be English
Output directoryCurrent project's docs/private/If user wants it elsewhere
Output formattxtIf user wants srt, vtt, json, or all

Step 3: Run transcription

TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 whisperx "<file_path>" \
  --diarize \
  --min_speakers <N> --max_speakers <N> \
  --language <lang> \
  --compute_type int8 \
  --output_dir "<output_dir>" \
  --output_format <format>

Important notes:

  • Always use --compute_type int8 on Mac (float16 is not supported on CPU)
  • Always set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 (PyTorch 2.6+ compatibility with pyannote)
  • First run downloads ~360MB model; subsequent runs are fast
  • Transcription of a 30-min file takes roughly 3-5 minutes on Mac

Step 4: Review output

After transcription completes:

  1. Read the output file
  2. Report: number of lines, speakers identified, any notable issues
  3. If speaker labels are generic (SPEAKER_00, SPEAKER_01), ask the user if they want to replace them with real names

Step 5: Optional cleanup

If the user wants cleaner output, offer to:

  • Replace speaker labels with real names (e.g., SPEAKER_00 → Bruce, SPEAKER_01 → Aaron)
  • Add timestamps (use --output_format all to get timestamped formats)

Step 6: Offer voice-briefing

After the transcript is ready, ask:

"Transcript is ready. Want me to run /voice-briefing on it to write it up as a business analysis in the context of everything else I know about your project and the market, and produce an mp3 reading of it and drop it in your dropbox folder?"

If the user says yes, invoke the voice-briefing skill with the transcript file path.

Troubleshooting

ErrorFix
float16 compute type not supportedUse --compute_type int8
UnpicklingError: Weights only load failedSet TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1
No --hf_token providedRun: ~/.local/pipx/venvs/whisperx/bin/huggingface-cli login --token <TOKEN>
401 Unauthorized for pyannoteAccept model terms on HuggingFace (links in Prerequisites)
whisperx: command not foundInstall: pipx install "whisperx @ git+https://github.com/m-bain/whisperX.git" --python python3.13

Without Speaker Diarization

If the user just wants plain transcription (no speaker identification), drop the diarize flags:

TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 whisperx "<file_path>" \
  --language <lang> \
  --compute_type int8 \
  --output_dir "<output_dir>" \
  --output_format <format>

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.