Faster Whisper Local

Local speech-to-text using faster-whisper. High-performance transcription with GPU acceleration support. Includes word-level timestamps and distilled models....

Security Scan

The skill's requirements, scripts, and instructions are coherent with a local faster-whisper transcription tool; it installs dependencies and runs only local transcription code and GPU-detection checks, with no unexpected credential access or external data exfiltration.

Purpose & Capability

Name/description (local faster-whisper STT) match the requested binaries (python3, ffmpeg), the Python dependencies (faster-whisper, torch), and the included scripts. Nothing requested (no cloud credentials, no unrelated binaries) is out of scope for a local transcription tool.

Instruction Scope

SKILL.md directs running setup.sh and scripts/transcribe.py to transcribe local audio files. The scripts check local system state (python version, ffmpeg, nvidia-smi, /proc/version for WSL) to detect GPUs — this is expected for enabling acceleration and is limited to local environment checks. There are no instructions to read unrelated user files, access secrets, or send transcription data to external endpoints.

Install Mechanism

This is instruction-only in the registry, but the package includes a setup.sh that creates a venv and pip-installs faster-whisper and PyTorch (from the official PyTorch index URL). Installing packages via pip and downloading models is expected but does execute third-party code from PyPI/torch wheels; this is normal for Python tools but carries the usual supply-chain risk. The index URL used for CUDA wheels is the official download.pytorch.org host.

Credentials

No environment variables, credentials, or config paths are requested. The script only needs local binaries (python3, ffmpeg) and will install Python packages — proportionate to its function.

Persistence & Privilege

always is false and the skill does not request persistent or elevated platform privileges or modify other skills. It creates a local venv and writes its own files, which is normal for a setup script.

Assessment

This skill appears to be what it says: a local transcription tool that sets up a Python virtual environment and installs faster-whisper and PyTorch. Before installing: (1) review and run the setup.sh in an isolated environment (container or dedicated machine) if you don't trust the source; (2) be aware pip will download and run third-party packages and the model weights (large download); (3) ensure ffmpeg and Python 3.10+ are installed from trusted system/package sources; (4) if you need extra assurance, create the venv manually and inspect requirements.txt and the scripts before running pip; (5) there's no evidence the skill exfiltrates data or asks for credentials, but transcriptions you run are processed locally — avoid running sensitive audio unless you control the host.

Faster-Whisper

High-performance local speech-to-text using faster-whisper.

Setup

1. Run Setup Script

Execute the setup script to create a virtual environment and install dependencies. It will automatically detect NVIDIA GPUs for CUDA acceleration.

./setup.sh

Requirements:

• Python 3.10 or later
• ffmpeg (installed on the system)

Usage

Use the transcription script to process audio files.

Basic Transcription

./scripts/transcribe audio.mp3

Advanced Options

• Specific Model: ./scripts/transcribe audio.mp3 --model large-v3-turbo
• Word Timestamps: ./scripts/transcribe audio.mp3 --word-timestamps
• JSON Output: ./scripts/transcribe audio.mp3 --json
• VAD (Silence Removal): ./scripts/transcribe audio.mp3 --vad

Available Models

• distil-large-v3 (default): Best balance of speed and accuracy.
• large-v3-turbo: Recommended for multilingual or highest accuracy tasks.
• medium.en, small.en: Faster, English-only versions.

Troubleshooting

• No GPU detected: Ensure NVIDIA drivers and CUDA are correctly installed. CPU transcription is significantly slower.
• OOM Error: Use a smaller model (e.g., small or base) or use --compute-type int8.