Local Vosk STT

Local speech-to-text using Vosk. Lightweight, fast, fully offline. Perfect for transcribing Telegram voice messages, audio files, or any speech-to-text task without cloud APIs.

Security Scan

The instructions claim a local Vosk transcription skill, but the SKILL.md references local scripts and runtime dependencies that are not present or declared — the skill is internally inconsistent and needs clarification before use.

Purpose & Capability

The description (local offline STT) matches the instructions (use vosk, download models). However SKILL.md instructs running ./skills/local-vosk/scripts/transcribe which implies bundled scripts/code that are not present in the package. Also the doc expects ffmpeg for decoding audio but the skill declares no required binaries. These gaps are disproportionate to the stated purpose.

Instruction Scope

Instructions tell the agent/user to run a local script path and to pip-install vosk and download models. Because there are no code files, an agent following these instructions would fail or attempt to run non-existent scripts. The instructions reference system actions (pip install, wget, unzip, writing to ~/vosk-models) that are reasonable for setup but include the unusual pip flag --break-system-packages without explanation.

Install Mechanism

There is no formal install spec (instruction-only), which is lower risk. The manual install commands point to a legitimate upstream site (alphacephei.com) for models and use pip/wget/unzip. Those sources are expected for Vosk models; no high-risk download URLs or shorteners are used. Still, because the skill lacks bundled code, it's unclear what the referenced scripts would do when present.

Credentials

The skill requests no environment variables or credentials, which is appropriate for an offline STT tool. No unrelated secrets are requested.

Persistence & Privilege

The skill does not request always:true and does not claim to modify other skills or system settings. It appears to be an on-demand instruction-only skill.

What to consider before installing

Don't install or run this skill as-is. SKILL.md expects a local script at ./skills/local-vosk/scripts/transcribe, but the package contains no code files — ask the publisher for the missing scripts or a corrected package. If you plan to run the provided setup commands yourself: ensure ffmpeg is installed (the README mentions it but the skill doesn't declare it), verify the model download source and checksums, and avoid running pip with unexplained flags like --break-system-packages unless you know what they do. Prefer a packaged release (includes the transcribe script) or run Vosk in an isolated environment/container until the skill's files and provenance are confirmed.

Lightweight local speech-to-text using Vosk. Fully offline after model download.

Use Cases

• Telegram voice messages — transcribe .ogg voice notes automatically
• Audio files — any format ffmpeg supports
• Offline transcription — no API keys, no cloud, no costs

Quick Start

# Transcribe Telegram voice message
./skills/local-vosk/scripts/transcribe voice_message.ogg

# Transcribe any audio
./skills/local-vosk/scripts/transcribe audio.mp3

# With language (default: en-us)
./skills/local-vosk/scripts/transcribe audio.wav --lang en-us

Supported Formats

Any format ffmpeg can decode: ogg (Telegram), mp3, wav, m4a, webm, flac, etc.

Models

Default model: vosk-model-small-en-us-0.15 (~40MB)

Other models available at https://alphacephei.com/vosk/models

Setup (if not installed)

pip3 install vosk --user --break-system-packages

# Download model
mkdir -p ~/vosk-models && cd ~/vosk-models
wget https://alphacephei.com/vosk/models/vosk-model-small-en-us-0.15.zip
unzip vosk-model-small-en-us-0.15.zip

Notes

• Quality is good for conversational speech
• For higher accuracy, use larger models or faster-whisper
• Processes audio at ~10x realtime on typical hardware
• Telegram voice messages are .ogg format — works out of the box