Deploy Silero VAD on Arm Ethos-U with ExecuTorch
Introduction
Prepare the environment and model inputs
Export Silero VAD for Ethos-U85
Build and run the Corstone-320 application
Validate the FVP speech probabilities
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Prepare to export the model

You prepared the model and two audio clips earlier. Now, use the calibration clip to quantize Silero VAD and the validation clip to create a host reference.

From the ExecuTorch repository root, activate the environment:

    

        
        
source .venv/bin/activate
source examples/arm/arm-scratch/setup_path.sh

Build the host quantized operators

The exporter requires the host shared library that registers the quantized operator out variants. Build the operator before exporting the model:

    

        
        
cmake \
  -S . \
  -B silero-vad-work/quantized_ops_aot \
  -DCMAKE_BUILD_TYPE=Release \
  -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON \
  -DEXECUTORCH_BUILD_KERNELS_QUANTIZED_AOT=ON \
  -DEXECUTORCH_BUILD_XNNPACK=OFF \
  -DPYTHON_EXECUTABLE="$(command -v python)"

cmake --build silero-vad-work/quantized_ops_aot \
  --target quantized_ops_aot_lib --parallel

export EXECUTORCH_QUANTIZED_OPS_AOT_LIBRARY="$(find \
  silero-vad-work/quantized_ops_aot/kernels/quantized \
  -name 'libquantized_ops_aot_lib.*' -type f -print -quit)"
test -s "$EXECUTORCH_QUANTIZED_OPS_AOT_LIBRARY"

Export the model

Run the exporter and save its output for the delegation check:

    

        
        
set -o pipefail

python3 examples/arm/silero_vad_example_ethos_u/model_export/export_silero_vad_ethos_u.py \
  --jit-model silero-vad-work/assets/silero_vad.jit \
  --calibration-audio silero-vad-work/assets/calibration.wav \
  --validation-audio silero-vad-work/assets/validation.wav \
  --output-path silero-vad-work/export/silero_vad_ethos_u.pte \
  --expected-output-path silero-vad-work/export/expected_probs.bin \
  --num-calibration-frames 32 \
  --num-validation-frames 0 2>&1 | \
  tee silero-vad-work/export/export.log

The output is similar to:

    

        
        Wrote expected probabilities to silero-vad-work/export/expected_probs.bin
Lowering to Ethos-U85...
Exported model saved to silero-vad-work/export/silero_vad_ethos_u.pte

Verify the generated outputs and the number of reference probabilities:

    

        
        
for artifact in \
  silero-vad-work/export/silero_vad_ethos_u.pte \
  silero-vad-work/export/expected_probs.bin \
  silero-vad-work/export/export.log; do
  test -s "$artifact" || {
    echo "Missing export artifact: $artifact" >&2
    exit 1
  }
done

test "$(wc -c < silero-vad-work/export/expected_probs.bin)" -eq $((79 * 4)) || {
  echo "Expected 79 float32 reference probabilities" >&2
  exit 1
}

test "$(grep -c 'CPU operators = 0' silero-vad-work/export/export.log)" -eq 2 || {
  echo "Expected both Vela subgraphs to contain no CPU operators" >&2
  exit 1
}

echo "Export artifacts and Ethos-U lowering verified."

Understand the streaming model

The application supplies one 512-sample audio frame at a time. The exported program keeps the LSTM hidden and cell state between calls. The program produces one speech probability every 32 ms.

Image Alt Text:Runtime diagram showing 64 context samples and a 512-sample frame entering the ExecuTorch program, model operations delegated to Ethos-U85, an internal int8 hidden and cell state reused between calls, and one speech probability emitted every 32 ms.Silero VAD streaming state and Ethos-U delegation boundary

The Vela output contains two subgraphs with no CPU operators: the main VAD network and a smaller state-quantization subgraph. Only small boundary conversions and the state update remain as portable ExecuTorch operations.

What you’ve accomplished and what’s next

You’ve exported Silero VAD as a stateful ExecuTorch program and saved the host reference output.

Next, you’ll build the bare-metal application and run it on the Corstone-320 FVP.

Back
Next