Silero VAD GGUF

GGUF weights of Silero VAD v5 for s2s.cpp, a C++17/GGML voice assistant. The model gives a speech probability for every 32 ms window of 16 kHz audio, its LSTM state carried from one window to the next.

Files

At this size a quant saves nothing and costs decisions, so F32 is the only variant.

Layout

convert.py builds the file from the ONNX export of onnx-community/silero-vad: the STFT basis is stored as a tensor, the two LSTM biases are summed, every dimension is a vad.* metadata key. The layout is the one src/silero.cpp reads through its C ABI, it does not load in llama.cpp.

Quick start

git clone https://github.com/ServeurpersoCom/s2s.cpp.git
cd s2s.cpp
git submodule update --init
./buildcuda.sh
./models.sh      # this file and the rest of the pipeline -> models/
./server.sh      # then open http://localhost:8088

s2s.cpp runs the VAD on the CPU with one thread: a GPU dispatch would cost more than the work.

Parity

Against onnxruntime on the upstream ONNX, over the example recording: probabilities at cosine similarity 1.000000000 (max abs 1.1e-6), speech decisions 539/539.

License

Upstream model: Silero VAD by Silero Team, MIT

GGUF tooling: s2s.cpp, MIT