anira-project/silero-vad

Silero VAD for anira

Silero VAD v6.2.3, a voice activity detector, for real-time inference with anira on ONNX Runtime, LiteRT, ExecuTorch (XNNPACK and Core ML) and LibTorch: 512 samples at 16 kHz in, the speech probability of the block out, its state as an explicit tensor.

Files

File	Engine	Size	SHA256
onnx/silero_vad_16k.onnx	ONNX Runtime	1,250,207	165af836d0a0442bfe5d5d91c3e715a995d55ac21ecc1179a96301fe140e7e2c
litert/silero_vad_16k.tflite	LiteRT, TensorFlow Lite	1,254,896	dfb2e98438dede7718e33cb3a63863b333b82e486ba122ebb803551094d488e9
executorch/silero_vad_16k.pte	ExecuTorch 1.3.1 (XNNPACK)	1,256,452	ea5aac11e1c8ba1e56675314fb74e593257cfd90646228a190b2a31ff05633c7
executorch/silero_vad_16k_coreml.pte	ExecuTorch 1.3.1 (Core ML, Apple)	1,323,745	a5f5df5e577c29b153a7a6e72c738ae49a95c6fc3d49d0305c4fa9ef897bbe99
executorch/silero_vad_16k_vulkan.pte	ExecuTorch 1.3.1 (Vulkan, GPUs)	1,255,684	5fca4378fb2b957b11df7755f82e1a1851d791405dd2fb19ebd47230e84d3b44
libtorch/silero_vad_16k.pt	LibTorch	1,249,106	5bfd137a62181f889343995da9bdde47bde3306e795a50553bf3a0b0a961adc2
upstream/silero_vad.onnx	upstream original	2,327,524	1a153a22f4509e292a94e67d6f9b85e8deb25b4988682b7e174c65279d8788e3

Inputs	Outputs
audio_in [1,1,512], state_in [1,320]	probability [1,1,1], state_out [1,320]

The state is the LSTM's h and c (128 each) and the last 64 samples of the previous call, which upstream's wrapper keeps outside the model and prepends to each block; start a stream with zeros. The LiteRT, ExecuTorch and LibTorch tensors are unnamed; their order is the ONNX model's.

Export

The scripts in export/ rebuild the 16 kHz model in eager PyTorch from the weights of upstream's TorchScript model (silero_vad.jit, same release) and export it. export.py checks the rebuild against upstream's TorchScript and ONNX models on a synthetic voiced signal (probabilities 0.001 to 0.96; largest deviation 7e-7) and writes ONNX and TorchScript; export_litert.py and export_executorch.py write LiteRT and ExecuTorch and check them against the ONNX export (1.3e-6); export_executorch.py --coreml lowers it to Core ML (silero_vad_16k_coreml.pte, Apple platforms, float32 on the CPU and GPU; 1.6e-6).

cd export
uv run --no-project export.py <silero_vad.jit> <silero_vad.onnx> <prefix>
uv run --no-project export_litert.py <silero_vad.jit> <prefix>
uv run --no-project export_executorch.py <silero_vad.jit> <prefix> [--coreml]

upstream/silero_vad.onnx is upstream's ONNX model, unchanged (its sr input selects 8 or 16 kHz; the caller passes the 64 samples of context).

Formats and checks

Every model here runs on ONNX Runtime, LiteRT and TensorFlow Lite, LibTorch, ExecuTorch (XNNPACK on the CPU, Core ML on Apple, Vulkan on GPUs) and as a native Core ML ML program, with the ONNX models' tensor names. anira foundation models streams every format of every description over twelve consecutive calls with the state fed back and checks that they agree (tools/check_streaming.py). The deviations are of the speech probability on the voiced test signal.

Format	Files	Max deviation from ONNX Runtime
ExecuTorch, Core ML (Apple, float32)	executorch/silero_vad_16k_coreml.pte	1.6e-6
ExecuTorch, Vulkan (GPUs)	executorch/silero_vad_16k_vulkan.pte	exported; not run here (no Vulkan runtime)
Core ML ML program (anira's native engine, coming with v3; macOS 13, iOS 16)	coreml/silero_vad_16k.mlpackage	1.6e-6

cd export
uv run --no-project export_executorch.py <silero_vad.jit> <prefix> [--coreml | --vulkan]
uv run --no-project export_coreml.py <silero_vad.jit> <prefix>

This repository holds the weights, the export scripts and the licence. The anira model description lives with the code in anira foundation models (models/speech-to-text), versioned with git. Pin a revision and check each file's SHA256 when you download:

https://huggingface.co/anira-project/silero-vad/resolve/<revision>/onnx/silero_vad_16k.onnx

Licence

MIT, © Silero Team: LICENSE, copied from snakers4/silero-vad @ v6.2.3, where the models are src/silero_vad/data/silero_vad.jit and silero_vad.onnx. The exports and the export scripts are MIT too.