gumblex / whisper_vad

Whisper.cpp Speech-to-text with Voice Acticity Detection

Whisper VAD

Whisper.cpp Speech-to-Text engine combined with Silero Voice Activity Detector. This improves transcription speed and quality, and can avoid hallucination of the model.

Run whisper_vad.py directly for transcribing any video/audio files into SRT subtitles, or import it as a library.

Dependencies

• ffmpeg (command)
• openblas (system library)
• cffi
• torch
• scipy
• zhconv: Chinese postprocess

Build and usage

1. pip install -r requirements.txt
2. make
3. python3 whisper_vad.py --help to see usage.

GPU

This currently only supports CLBlast and AMD HIPBLAS.

CLBlast

Dependencies: libclblast, OpenCL

Build: WHISPER_CLBLAST=1 make

HIPBLAS

Dependencies: libhipblas, libamdhip64, librocblas

Build:

1. Build original Whisper.cpp
2. Copy ggml-cuda.o to whisper_cpp
3. WHISPER_HIPBLAS=1 make