← Back to Skills Marketplace
1397
Downloads
0
Stars
12
Active Installs
9
Versions
Install in OpenClaw
/install faster-whisper-local-service
Description
OpenClaw local speech-to-text backend using faster-whisper over HTTP on 127.0.0.1:18790. Use when you want voice transcription without external APIs, without...
README (SKILL.md)
Faster Whisper Local Service
Provision a local STT backend used by voice skills.
What this sets up
- Python venv for faster-whisper
transcribe-server.pyHTTP endpoint athttp://127.0.0.1:18790/transcribe- systemd user service:
openclaw-transcribe.service
Important: Model download on first run
On first startup, faster-whisper downloads model weights from Hugging Face (~1.5 GB for medium). This requires internet access and disk space. After the initial download, models are cached locally and the service runs fully offline.
| Model | Download size | RAM usage |
|---|---|---|
| tiny | ~75 MB | ~400 MB |
| base | ~150 MB | ~500 MB |
| small | ~500 MB | ~800 MB |
| medium | ~1.5 GB | ~1.4 GB |
| large-v3 | ~3.0 GB | ~3.5 GB |
To pre-download models in an air-gapped environment, see faster-whisper docs.
Security notes
Network isolation
- Binds to
127.0.0.1only — not reachable from the network. - CORS restricted to a single origin (
https://127.0.0.1:8443by default). - No credentials, API keys, or secrets are used or stored.
Input validation
- Upload size limit: Requests exceeding the configured limit are rejected before processing (HTTP 413). Default: 50 MB, configurable via
MAX_UPLOAD_MB. - Magic-byte check: Only files with recognized audio signatures (WAV, OGG, FLAC, MP3, WebM, M4A) are accepted. Unrecognized formats are rejected (HTTP 415) before reaching GStreamer.
- Subprocess safety: All arguments to
gst-launch-1.0are passed as a list — no shell expansion or injection is possible.
GStreamer dependency
The service uses GStreamer's decodebin for audio format conversion. Like any media library, GStreamer's parsers process binary data and should be kept up to date. Mitigation: install gst-launch-1.0 from your OS vendor's trusted packages and apply security updates regularly. The magic-byte pre-filter above reduces the attack surface by rejecting non-audio payloads before they reach GStreamer.
No data exfiltration
- No outbound network calls (after initial model download).
- No telemetry, analytics, or phone-home behavior.
- Temporary files are created in a per-request
TemporaryDirectoryand cleaned up immediately.
Reproducibility defaults
- Pinned package:
faster-whisper==1.1.1(override via env) - Explicit dependency check for
gst-launch-1.0 - CORS restricted to one origin by default
- Configurable workspace/service paths (no hardcoded user path)
Deploy
bash scripts/deploy.sh
With custom settings:
WORKSPACE=~/.openclaw/workspace \
TRANSCRIBE_PORT=18790 \
WHISPER_MODEL_SIZE=medium \
WHISPER_LANGUAGE=auto \
TRANSCRIBE_ALLOWED_ORIGIN=https://10.0.0.42:8443 \
bash scripts/deploy.sh
Language setting
Default: auto (auto-detect language). Set WHISPER_LANGUAGE=de for German-only, en for English-only, etc. Fixed language is faster and more accurate if you only use one language.
Idempotent: safe to run repeatedly.
What this skill modifies
| What | Path | Action |
|---|---|---|
| Python venv | $WORKSPACE/.venv-faster-whisper/ |
Creates venv, installs faster-whisper via pip |
| Transcribe server | $WORKSPACE/voice-input/transcribe-server.py |
Writes server script |
| Systemd service | ~/.config/systemd/user/openclaw-transcribe.service |
Creates + enables persistent service |
| Model cache | ~/.cache/huggingface/ |
Downloads model weights on first run |
Uninstall
systemctl --user stop openclaw-transcribe.service
systemctl --user disable openclaw-transcribe.service
rm -f ~/.config/systemd/user/openclaw-transcribe.service
systemctl --user daemon-reload
Optional full cleanup:
rm -rf ~/.openclaw/workspace/.venv-faster-whisper
rm -f ~/.openclaw/workspace/voice-input/transcribe-server.py
Verify
bash scripts/status.sh
Expected:
- service
active - endpoint responds (HTTP 200/500 acceptable for invalid sample payload)
Notes
- This skill provides backend transcription only.
- Pair with
webchat-voice-proxyfor browser mic + HTTPS/WSS integration. - For one-step install, use
webchat-voice-full-stack(deploys backend + proxy in order).
Usage Guidance
This appears to be a legitimate local transcription installer. Before installing, consider: (1) the deploy script will pip-install faster-whisper into a user venv and create a systemd user service (it runs with your user privileges); (2) on first run faster-whisper will download large model weights from Hugging Face — ensure you want that network activity and disk use; (3) the service uses gst-launch-1.0 (OS package) for audio conversion — keep GStreamer updated; (4) review transcribe-server.py (included) if you want to audit behavior yourself; and (5) if you need stronger isolation, run the service in a dedicated user account, container, or VM. If you do not trust the faster-whisper package source, consider inspecting or pinning dependencies before installing.
Capability Analysis
Type: OpenClaw Skill
Name: faster-whisper-local-service
Version: 0.2.0
The skill provisions a local speech-to-text service using the `faster-whisper` library. It includes a deployment script (`scripts/deploy.sh`) that sets up a Python virtual environment, a transcription server (`transcribe-server.py`), and a systemd user service. The implementation includes security best practices such as magic-byte verification for audio files, upload size limits, and safe subprocess execution for GStreamer. No evidence of data exfiltration, malicious persistence, or prompt injection was found.
Capability Assessment
Purpose & Capability
Name/description match the behavior: the scripts create a venv, install faster-whisper, write a local HTTP server, and register a per-user systemd service. All created files and env vars are clearly for configuring a local STT backend.
Instruction Scope
SKILL.md and scripts only perform actions needed for a local transcription service: check for python and gst-launch, create venv, pip-install faster-whisper, write server script and systemd unit, and run the service. The server enforces upload limits, magic-byte checks, and binds to 127.0.0.1. It does spawn gst-launch for audio conversion (via subprocess with argument list), which is expected for media handling and is explicitly called out and constrained.
Install Mechanism
No registry install spec, but deploy.sh pip-installs faster-whisper from PyPI into a local venv (moderate risk but expected). Model weights are downloaded from Hugging Face on first run (large files, requires Internet). No downloads from untrusted custom URLs or URL shorteners.
Credentials
The skill does not require secrets or unrelated environment variables. Environment variables present are configuration knobs (port, model size, device, allowed origin, max upload) and are proportional to the stated function.
Persistence & Privilege
The skill creates a per-user systemd service (~~/.config/systemd/user/...) and files under a configurable workspace in the user's home. It does not request always:true, system-wide privileges, or modify other skills' configurations.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install faster-whisper-local-service - After installation, invoke the skill by name or use
/faster-whisper-local-service - Provide required inputs per the skill's parameter spec and get structured output
Version History
v0.2.0
Security hardening: input validation (magic-byte check, configurable upload size limit), GStreamer subprocess timeout, sanitized error responses, restructured security docs
v0.1.7
Refresh tags for local/offline STT discoverability.
v0.1.6
Fixed misleading 'fully offline' claim: now documents model download on first run with size table. Made language configurable (WHISPER_LANGUAGE, default: auto instead of hardcoded 'de'). Added security notes about gst-launch-1.0 and untrusted audio. Added system modifications table and uninstall instructions.
v0.1.5
Improved discoverability: expanded description with STT, speech-to-text, transcription, microphone, voice input keywords
v0.1.4
Ranking tune: stronger OpenClaw/local STT naming and discoverability keywords.
v0.1.3
Searchability update: explicit OpenClaw/local/offline/no extra API cost wording in description.
v0.1.2
Added reference to webchat-voice-full-stack for one-step backend+proxy deployment.
v0.1.1
Hardened deploy: configurable paths, dependency checks, pinned faster-whisper version, and restrictive configurable CORS origin.
v0.1.0
Initial release: installs local faster-whisper transcription service with systemd user unit and health checks.
Metadata
Frequently Asked Questions
What is Faster Whisper Local Service?
OpenClaw local speech-to-text backend using faster-whisper over HTTP on 127.0.0.1:18790. Use when you want voice transcription without external APIs, without... It is an AI Agent Skill for Claude Code / OpenClaw, with 1397 downloads so far.
How do I install Faster Whisper Local Service?
Run "/install faster-whisper-local-service" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Faster Whisper Local Service free?
Yes, Faster Whisper Local Service is completely free (open-source). You can download, install and use it at no cost.
Which platforms does Faster Whisper Local Service support?
Faster Whisper Local Service is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Faster Whisper Local Service?
It is built and maintained by neldar (@neldar); the current version is v0.2.0.
More Skills