Whisper STT

Free local speech-to-text transcription using OpenAI Whisper. Transcribe audio files (mp3, wav, m4a, ogg, etc.) to text without API costs. Use when: (1) User needs audio/video transcription, (2) Converting voice memos to text, (3) Generating subtitles (SRT/VTT), (4) Free local STT without cloud API costs.

Whisper STT Skill

Free, local speech-to-text using OpenAI Whisper.

Prerequisites

Install dependencies (one-time setup):

bash

pip install openai-whisper torch

Optional: Install ffmpeg for broader format support:

• macOS: brew install ffmpeg
• Ubuntu: sudo apt install ffmpeg

Usage

Transcribe an audio file

bash

python ~/.openclaw/skills/whisper-stt/scripts/transcribe.py <audio_file>

Options

Option	Description
--model	Model size: tiny, base, small, medium, large, large-v3-turbo (default: base)
--language, -l	Language code: zh, en, ja, etc. (auto-detect if not specified)
--output, -o	Output format: json, txt, srt, vtt (default: json)

Examples

Chinese audio to text:

bash

python ~/.openclaw/skills/whisper-stt/scripts/transcribe.py recording.m4a --language zh --output txt

Generate subtitles (SRT):

bash

python ~/.openclaw/skills/whisper-stt/scripts/transcribe.py video.mp4 --output srt > subtitles.srt

Use faster model:

bash

python ~/.openclaw/skills/whisper-stt/scripts/transcribe.py audio.mp3 --model tiny --output txt

High accuracy (slower):

bash

python ~/.openclaw/skills/whisper-stt/scripts/transcribe.py audio.mp3 --model large-v3 --output txt

Model Selection Guide

Model	Speed	Accuracy	VRAM/RAM	Best For
tiny	~32x	Basic	~1GB	Quick tests, low resource
base	~16x	Good	~1GB	Balanced speed/accuracy
small	~6x	Better	~2GB	Better accuracy
medium	~2x	Very Good	~5GB	High accuracy
large	1x	Excellent	~10GB	Best quality
large-v3-turbo	~8x	Excellent	~6GB	Fast + accurate (recommended)

Troubleshooting

"ModuleNotFoundError: No module named 'whisper'" → Run: pip install openai-whisper torch

"ffmpeg not found" → Install ffmpeg or convert audio to WAV format first

Slow transcription → Use smaller model (tiny/base) or ensure GPU is available (Apple Silicon MPS, NVIDIA CUDA)

Poor accuracy on Chinese → Use --language zh explicitly and consider larger model (medium/large)

Output Formats

• json: Full result with segments, timestamps, and metadata
• txt: Plain text transcription only
• srt: SubRip subtitle format with timing
• vtt: WebVTT subtitle format for web players

Credits

Powered by OpenAI Whisper - open source speech recognition.