Audio Transcriber
Transcribe any audio or video file to text using Whisper — runs locally, free, no API keys.
Requirements (Already Installed)
faster-whisper (Python package) — the transcription engineffmpeg (at ~/bin/ffmpeg) — audio processingtranscribe.py (at this skill's location) — the transcription script
How to Transcribe
Run the transcription script using Bash:
python3 ~/.claude/skills/audio-transcriber/transcribe.py "<path_to_audio_file>"
Options
| Flag |
Values |
Default |
What it does |
--model |
tiny, base, small, medium, large-v3 |
base |
Larger = more accurate but slower |
--output |
text, srt, vtt, segments, all |
text |
Output format |
--outfile |
file path |
(prints to screen) |
Save to file instead of printing |
Model Selection Guide
| Model |
Speed |
Accuracy |
Best for |
tiny |
Fastest (~10x realtime) |
Good for clear speech |
Quick previews, clear recordings |
base |
Fast (~5x realtime) |
Good |
Default — works well for most files |
small |
Moderate (~2x realtime) |
Very good |
Accented speech, background noise |
medium |
Slow (~1x realtime) |
Excellent |
Important transcripts, multiple speakers |
large-v3 |
Slowest |
Best available |
Critical accuracy, foreign languages |
First run of each model size downloads the model (~50MB-3GB). Subsequent runs are instant.
Examples
Basic transcription (print to screen):
python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/recording.mp3"
Save to text file:
python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/recording.mp3" --outfile "/path/to/transcript.txt"
Higher accuracy with medium model:
python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/recording.mp3" --model medium
Generate subtitles (SRT format):
python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/video.mp4" --output srt --outfile "/path/to/subtitles.srt"
All formats at once (text + SRT + VTT + segments):
python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/audio.m4a" --output all --outfile "/path/to/output.txt"
Supported File Types
Any format ffmpeg can read: .mp3, .wav, .m4a, .mp4, .flac, .ogg, .webm, .aac, .wma, .mov, .avi, .mkv, and many more.
Workflow
- User provides an audio/video file path
- Choose the right model —
base for most files, medium or large-v3 for noisy/accented/critical content
- Run the transcription script using Bash tool
- Read the output — either from the Bash output or from the saved file
- Post-process if needed — clean up, summarize, extract key points, etc.
Tips
- For long files (60+ minutes), use
--model base first for a quick draft, then re-run specific sections with medium if needed
- The first run of a new model size will download it (~30 seconds to a few minutes). After that, it's cached.
- Output SRT/VTT formats if the user wants timestamps with each segment
- The
segments output shows timestamped chunks — useful for finding specific moments
- If transcription quality is poor, try upgrading the model size one step
- The script auto-detects language — works with English, Spanish, French, and 90+ other languages
1---2name: audio-transcriber3description: Transcribe audio and video files to text using Whisper (local, free, no API key needed). Use this skill whenever the user asks to transcribe, convert audio to text, get a transcript, subtitle, or caption any audio or video file. Supports mp3, wav, m4a, mp4, flac, ogg, webm, and more. Trigger for "transcribe", "transcript", "audio to text", "what does this audio say", "convert this recording", "speech to text", "subtitles", "captions", "listen to this file", or any request involving extracting text from audio or video files.4---56# Audio Transcriber78Transcribe any audio or video file to text using Whisper — runs locally, free, no API keys.910## Requirements (Already Installed)1112- `faster-whisper` (Python package) — the transcription engine13- `ffmpeg` (at `~/bin/ffmpeg`) — audio processing14- `transcribe.py` (at this skill's location) — the transcription script1516## How to Transcribe1718Run the transcription script using Bash:1920```bash21python3 ~/.claude/skills/audio-transcriber/transcribe.py "<path_to_audio_file>"22```2324### Options2526| Flag | Values | Default | What it does |27|------|--------|---------|-------------|28| `--model` | `tiny`, `base`, `small`, `medium`, `large-v3` | `base` | Larger = more accurate but slower |29| `--output` | `text`, `srt`, `vtt`, `segments`, `all` | `text` | Output format |30| `--outfile` | file path | (prints to screen) | Save to file instead of printing |3132### Model Selection Guide3334| Model | Speed | Accuracy | Best for |35|-------|-------|----------|----------|36| `tiny` | Fastest (~10x realtime) | Good for clear speech | Quick previews, clear recordings |37| `base` | Fast (~5x realtime) | Good | Default — works well for most files |38| `small` | Moderate (~2x realtime) | Very good | Accented speech, background noise |39| `medium` | Slow (~1x realtime) | Excellent | Important transcripts, multiple speakers |40| `large-v3` | Slowest | Best available | Critical accuracy, foreign languages |4142**First run of each model size downloads the model (~50MB-3GB). Subsequent runs are instant.**4344### Examples4546**Basic transcription (print to screen):**47```bash48python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/recording.mp3"49```5051**Save to text file:**52```bash53python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/recording.mp3" --outfile "/path/to/transcript.txt"54```5556**Higher accuracy with medium model:**57```bash58python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/recording.mp3" --model medium59```6061**Generate subtitles (SRT format):**62```bash63python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/video.mp4" --output srt --outfile "/path/to/subtitles.srt"64```6566**All formats at once (text + SRT + VTT + segments):**67```bash68python3 ~/.claude/skills/audio-transcriber/transcribe.py "/path/to/audio.m4a" --output all --outfile "/path/to/output.txt"69```7071## Supported File Types7273Any format ffmpeg can read: `.mp3`, `.wav`, `.m4a`, `.mp4`, `.flac`, `.ogg`, `.webm`, `.aac`, `.wma`, `.mov`, `.avi`, `.mkv`, and many more.7475## Workflow76771. **User provides an audio/video file path**782. **Choose the right model** — `base` for most files, `medium` or `large-v3` for noisy/accented/critical content793. **Run the transcription script** using Bash tool804. **Read the output** — either from the Bash output or from the saved file815. **Post-process if needed** — clean up, summarize, extract key points, etc.8283## Tips8485- For **long files** (60+ minutes), use `--model base` first for a quick draft, then re-run specific sections with `medium` if needed86- The **first run** of a new model size will download it (~30 seconds to a few minutes). After that, it's cached.87- Output **SRT/VTT** formats if the user wants timestamps with each segment88- The `segments` output shows timestamped chunks — useful for finding specific moments89- If transcription quality is poor, try upgrading the model size one step90- The script auto-detects language — works with English, Spanish, French, and 90+ other languages
prayceo