Whisper.cpp Local Transcription Engine
Runs OpenAI Whisper models locally via whisper.cpp with GGML quantized weights for CPU-efficient transcription. Supports beam search decoding, VAD-based segmentation, and SRT/VTT subtitle output formats.
npx skills add agentskillexchange/skills --skill whisper-cpp-local-transcription-engine
This skill provides fully local speech-to-text transcription using whisper.cpp, the C/C++ port of OpenAI’s Whisper model. It downloads GGML-quantized model weights (tiny, base, small, medium, large-v3) and runs inference on CPU with optional GPU acceleration via CUDA or Metal backends. The core command runs: ./main -m models/ggml-large-v3.bin -f input.wav -osrt -ocsv –beam-size 5 –best-of 5. The skill handles audio preprocessing by converting input files to 16kHz mono WAV using FFmpeg: ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav. It supports multiple output formats: plain text (-otxt), SRT subtitles (-osrt), VTT subtitles (-ovtt), CSV with timestamps (-ocsv), and JSON with word-level timing (-ojson). Advanced features include voice activity detection (–no-speech-thold 0.6) to skip silence, language auto-detection (–language auto), and translation mode (–translate) for non-English audio. The skill implements batch processing for multiple files with progress tracking. Performance tuning includes thread count configuration (–threads N), and the skill benchmarks inference speed as a ratio of audio duration. For long recordings, it uses –max-len 0 to process without segment length limits.