← Back to Skills Marketplace
8859
Downloads
5
Stars
0
Active Installs
20
Versions
Install in OpenClaw
/install faster-whisper
Description
Local speech-to-text using faster-whisper. 4-6x faster than OpenAI Whisper with identical accuracy; GPU acceleration enables ~20x realtime transcription. SRT...
Usage Guidance
Install only if you are comfortable with a transcription tool that can contact external media sites, Hugging Face, package indexes, and optional Docker images. Avoid using it with private RSS/URLs or sensitive Hugging Face tokens unless you can isolate the environment, prefer pinned dependencies/images, and do not open generated HTML transcripts from untrusted audio without sanitizing them first.
Capability Analysis
Type: OpenClaw Skill
Name: faster-whisper
Version: 1.5.1
This skill is classified as suspicious due to its broad capabilities, which include downloading content from arbitrary URLs via `yt-dlp`, executing `ffmpeg` for audio/video processing and subtitle burning, and performing self-updates of its core dependency. While these actions are plausibly aligned with the stated purpose of a comprehensive transcription tool, they grant significant access to the network and local file system. The `SKILL.md` agent guidance does not contain any malicious prompt injection attempts, and the `setup.sh` and `scripts/transcribe.py` files implement these powerful features using `subprocess.run()` with argument lists, which mitigates direct shell injection vulnerabilities. However, the inherent power of these operations, even when used for legitimate purposes, elevates the risk profile beyond a purely benign classification.
Capability Assessment
Credentials
Local file reads/writes, ffmpeg and yt-dlp subprocesses, remote media retrieval, model cache access, and optional Hugging Face token access fit the transcription purpose but are broad and not strongly contained, especially for arbitrary URLs and RSS enclosure URLs.
Install Mechanism
Setup creates a local virtual environment and installs expected ML dependencies, but dependencies are lower-bound or unpinned, optional packages are unconstrained, update paths pull latest packages, and the Docker example uses a mutable latest-cuda image tag.
Instruction Scope
The skill often tells the agent to add risky flags only when explicitly requested, but it lacks strong confirmation or destination-scope rules for URL/RSS downloads, Hugging Face token use, in-place updates, and server-mode examples.
Persistence & Privilege
No backdoor, scheduled task, or hidden service persistence was found; persistence is mainly expected caches, output files, a skill venv, optional kept temp files, and user-invoked dependency updates.
Purpose & Capability
The core speech-to-text purpose is coherent, and URL/RSS transcription, diarization, subtitle output, and batch processing are disclosed, but the local/offline framing under-discloses network fetching, model downloads, credential use, and generated active HTML output.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install faster-whisper - After installation, invoke the skill by name or use
/faster-whisper - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.5.1
- Fixed --skip-existing in multi-format mode to check ALL format outputs before skipping
- Fixed --no-timestamps conflict check missing lrc, ass, ttml formats
- Fixed --speaker-names silently doing nothing without --diarize; now prints a warning
- Batch summary now shows skipped file count when --skip-existing is active
v1.5.0
- docs: update default model from distil-large-v3 to distil-large-v3.5
- fix: setup.sh --check hangfix + skill.json ffmpeg optional
- fix(transcribe): clean-filler word list, fuzzy search tokens, URL temp cleanup
- fix(multi-format): create output dir in single-file mode
- feat: add CSV output, language-map, batch ETA estimate
- feat: add TTML output, transcript search, chapter detection, speaker audio export
- feat: distil auto-condition, log-level, ffmpeg clarification
- fix: rename --without-timestamps to --no-timestamps
- feat: add 17 new features — upstream params + LRC/detect-language/merge-sentences/stats/stdin/template
v1.4.5
- Fix author field to match GitHub username (ThePlasmak)
v1.4.4
- Declare yt-dlp and HuggingFace token as optional dependencies in skill.json
- Sync SKILL.md frontmatter version and author with skill.json
v1.4.3
- Auto-run wav2vec2 alignment whenever word timestamps are computed
- Remove --precise flag (alignment is automatic, flag kept as hidden compat alias)
- Alignment triggers for --word-timestamps, --diarize, --min-confidence
- No overhead for basic transcription (fast path unchanged)
v1.4.1
- Add --precise flag for wav2vec2 forced alignment (~10ms word accuracy)
- Uses torchaudio MMS model (multilingual, cached for batch processing)
- Runs before diarization when combined (improves speaker assignment)
- Install torchaudio alongside torch in setup.sh
v1.3.0
- Add SRT and VTT subtitle output formats (--format srt/vtt)
- Add speaker diarization via pyannote.audio (--diarize) with word-level accuracy
- Add URL/YouTube input with auto yt-dlp download
- Add batch processing with glob patterns, directories, and --skip-existing
- Add initial prompt support for domain terminology (--initial-prompt)
- Add confidence-based segment filtering (--min-confidence)
- Add performance stats after each transcription (duration, realtime factor)
- Unify output under --format flag (text/json/srt/vtt), keep --json for backward compat
- Add agent guidance for minimal invocation (don't load unused features)
v1.2.0
- Default model changed to distil-large-v3.5 (lower WER: 7.08 vs 7.53, same speed as v3)
- Trained on 4x more data (98k hours) with improved robustness
v1.1.0
- Use BatchedInferencePipeline by default (~3x faster; 69s → 23s on 21-min file with distil-large-v3)
- VAD enabled by default in batched mode
- Add --batch-size option (default: 8; reduce if OOM)
- Add --no-batch flag to fall back to standard WhisperModel
- Add --hotwords support for boosting recognition of specific terms
- Bump tested version: faster-whisper 1.2.1
v1.0.12
- Fix skill title display on ClawdHub
v1.0.11
- Prefer distil-large-v3 over large-v3-turbo as the recommended model
v1.0.9
- docs: rebrand from Moltbot/MoltHub to OpenClaw/ClawHub
v1.0.7
- Removed Windows-native references from SKILL.md (setup.ps1, transcribe.cmd, winget) since ClawHub cannot distribute .ps1/.cmd files
- Windows users should use WSL2 or get Windows scripts from the GitHub repo directly
v1.0.6
- Added .clawdhubignore to exclude README.md, CHANGELOG.md, LICENSE from published package
- Fixed requires.bins in skill.json (python3, ffmpeg)
- Added platforms field to skill.json
- Updated metadata key from moltbot to openclaw in SKILL.md
v1.0.5
Fix metadata: add requires.bins to skill.json, add platforms, update moltbot to openclaw in SKILL.md
v1.0.4
- Fixed skill title and metadata
- Removed development files from published package
v1.0.3
Fix skill title (was 'Faster Whisper Clean' due to temp file naming)
v1.0.2
- Improve skill discovery, error handling, some copyediting
- Add skill.json
- Edit README to reduce confusion as Moltbot may refer to the service
v1.0.1
Remove install metadata (as ClawdHub's install section is confusing); add python3 to required binaries
v1.0.0
Initial public release of faster-whisper.
- Local speech-to-text using faster-whisper (CTranslate2 backend), ~4-6x faster than OpenAI Whisper, with identical accuracy.
- Supports GPU acceleration for ~20x realtime transcription; automatic hardware detection and setup for Windows.
- Offers both standard and distilled models, with selectable accuracy/speed tradeoffs and word-level timestamps.
- Cross-platform: Windows (including WSL2), Linux, and macOS (Apple Silicon supported).
- Setup scripts provided for all platforms, including automatic installation of dependencies and GPU support where possible.
- Includes extensive usage documentation, quick-start commands, model selection guide, and troubleshooting tips.
Metadata
Frequently Asked Questions
What is Faster Whisper?
Local speech-to-text using faster-whisper. 4-6x faster than OpenAI Whisper with identical accuracy; GPU acceleration enables ~20x realtime transcription. SRT... It is an AI Agent Skill for Claude Code / OpenClaw, with 8859 downloads so far.
How do I install Faster Whisper?
Run "/install faster-whisper" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Faster Whisper free?
Yes, Faster Whisper is completely free (open-source). You can download, install and use it at no cost.
Which platforms does Faster Whisper support?
Faster Whisper is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Faster Whisper?
It is built and maintained by Sarah Mak (@theplasmak); the current version is v1.5.1.
More Skills