Silero VAD C Port

Standalone C port of the 16 kHz Silero VAD model with embedded weights and no ONNX dependency.

This repository packages the model as a small shared library for: Windows, macOS, and Linux.

The code currently focuses on the 16 kHz model path and supports both chunked inference and full-audio probability extraction.

Current status

• 16 kHz model path implemented
• Embedded-weight builds supported
• Windows DLL, Linux and macOS shared-library builds supported through CMake
• Several performance optimizations are already implemented:
  • stft-conv replaced by fixed 256-point FFT
  • fused STFT to first conv path
  • shared SIMD abstraction for scalar, SSE, AVX2 and NEON backends
  • AVX2 conv and LSTM kernels for x86_64
  • SSE conv and LSTM kernels for x86 and x86_64
  • NEON conv and LSTM kernels for arm64

Python example

Load a built shared library and run full-audio inference:

from run_silero_vad_clib import SileroVadClib
from src.silero_vad.utils_vad import read_audio

audio = read_audio("src/silero_vad/test/tests_data_test.wav", 16000)

with SileroVadClib("./downloads/silero-vad-linux-x86_64-avx2/silero_vad.so") as model:
    probs = model.forward_audio(audio)

Download binaries

Download prebuilt binaries from the releases site, or by running download_releases.py.

This command downloads the binaries for the local system and unzips them.

python download_releases.py --unzip

Use --os all to download all binaries, independent of the system.

python download_releases.py --os all --unzip

Benchmark

Full-audio benchmark on the 60-second test clip at src/silero_vad/test/tests_data_test.wav.

CPU	Build	t (s)	t/d
AMD Ryzen 9 7900	silero-vad-windows-x64-sse	0.068892	0.001148
AMD Ryzen 9 7900	silero-vad-windows-x64-avx2	0.058982	0.000983
AMD Ryzen 9 7900	torch.hub(onnx=False)	0.447120	0.007452
AMD Ryzen 9 7900	torch.hub(onnx=True)	0.257800	0.004297

t is the inference time averaged over 10 iterations. d is the audio duration (here 60 seconds). Both native C builds are much faster than the torch.hub baselines on the tested machine. The AVX2 build is the fastest result here, with the SSE build close behind, and both run at well under 1% of real time for this 60-second input.

Repository layout

• silero_vad.c / silero_vad.h / silero_vad_simd.h
  • core C implementation, public API, and SIMD abstraction layer
• export_silero_vad_weights.py
  • exports weights from safetensors
• export_silero_vad_jit_weights.py
  • exports weights from the Torch hub JIT model
• run_silero_vad_clib.py
  • Python ctypes example / test runner for the shared library
• package_release.py
  • creates release folders and zip files from built binaries

Model weights

Two weight sources are supported:

• jit
  • exported from torch.hub.load(..., model='silero_vad')
  • best choice if you want parity with the original Silero Torch hub model
• safetensors
  • exported from the local safetensors checkpoint

Building and packaging

Build and release commands are collected in:

• building.md

That file includes: Windows builds, Linux builds, macOS builds, and release packaging commands

Public API

The public API is declared in silero_vad.h.

Main entry points:

• silero_vad_model_create
• silero_vad_model_destroy
• silero_vad_model_reset
• silero_vad_model_forward
• silero_vad_model_forward_audio

The current full-model path is designed around 16 kHz audio and typically uses:

• 576 samples per chunk
• 64 samples of left context
• 512 new samples

Notes

• x86 and x86_64 SIMD backends are now build-selectable:
  • scalar baseline
  • SSE with SILERO_VAD_ENABLE_SSE=ON
  • AVX2 with SILERO_VAD_ENABLE_AVX2=ON
• Use the baseline or SSE build as the compatibility build for older x86 and x86_64 CPUs without AVX2.
• Use the AVX2 build when you want the faster x86_64 path on AVX2-capable CPUs.
• SILERO_VAD_FAST_MATH is available, but on current tests it did not improve performance.
• ARM builds support the NEON-optimized path when SILERO_VAD_ENABLE_NEON=ON.

License

This repository is based on the MIT-licensed Silero VAD project.