Prepare the environment and model inputs

Understand the workflow

Voice activity detection (VAD) classifies short audio frames as speech or silence. VAD is useful for voice assistants, transcription pipelines, and other systems that should avoid processing silent audio.

You’ll deploy the 16 kHz Silero VAD model with ExecuTorch. The workflow quantizes the model, lowers supported operations to the Arm Ethos-U backend, builds a bare-metal application, and runs it on a Corstone-320 Fixed Virtual Platform (FVP). You don’t need a physical development board.

The application processes 512 audio samples every 32 ms. It keeps the long short-term memory (LSTM) hidden and cell state inside the ExecuTorch program between frames, then produces one speech probability for each frame.

The host uses the validation clip to generate reference probabilities. The bare-metal application processes the same clip on the FVP. The final comparison verifies that both paths produce the same speech decisions.

Silero VAD workflow separated into prepare, run, and verify lanes

Use the FVP for functional validation. The FVP’s Ethos-U model is cycle accurate, but don’t use its Cortex-M CPU model for CPU performance measurements.

Check your development machine

Use Python 3.12 with development headers and virtual-environment support. The release wheels require glibc 2.28 or later on Linux, or macOS 15 or later on Apple silicon. On Linux, the FVP also requires libstdc++.so.6 providing GLIBCXX_3.4.26 or later. Run the following preflight check before downloading the source; CMake is installed in the virtual environment later:

    

        
        
case "$(uname -s)/$(uname -m)" in
  Linux/x86_64|Linux/aarch64|Linux/arm64|Darwin/arm64)
    echo "Supported host: $(uname -s)/$(uname -m)"
    ;;
  *)
    echo "Unsupported host: $(uname -s)/$(uname -m)" >&2
    exit 1
    ;;
esac

for tool in python3.12 git c++ curl; do
  command -v "$tool" >/dev/null || {
    echo "Missing required tool: $tool" >&2
    exit 1
  }
done

if ! command -v ninja >/dev/null && ! command -v make >/dev/null; then
  echo "Install Ninja or Make before continuing." >&2
  exit 1
fi

printf 'int main() { return 0; }\n' | \
  c++ -std=c++17 -x c++ -fsyntax-only -

python3.12 --version

    

Create an isolated ExecuTorch environment

Use the source revision from the Silero VAD end-to-end example with the released ExecuTorch 1.5.1 Python package:

    

        
        
git clone https://github.com/pytorch/executorch.git
cd executorch
git fetch origin d78bc1b0ba1c33a90d2d93662de118577f37efd4
git checkout d78bc1b0ba1c33a90d2d93662de118577f37efd4
git submodule sync --recursive
git submodule update --init --recursive

    

Confirm the checked-out revision:

    

        
        
git rev-parse --short=12 HEAD

    

The expected output is:

    

        
        d78bc1b0ba1c

        
    

Create and activate a Python virtual environment:

    

        
        
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip

    

Install the Arm backend tools

The Arm setup script installs the Arm GNU Toolchain and, on Linux, the Corstone FVPs. Review the license terms presented by the script before using its EULA acceptance option.

    

        
        
mkdir -p examples/arm/arm-scratch "$HOME/.armlm"
git clone https://github.com/Arm-Examples/FVPs-on-Mac.git \
  examples/arm/arm-scratch/FVPs-on-Mac
git -C examples/arm/arm-scratch/FVPs-on-Mac checkout --detach \
  cc7a4be5456db085619647811fac29e9795f806e
bash examples/arm/arm-scratch/FVPs-on-Mac/build.sh
export PATH="$PWD/examples/arm/arm-scratch/FVPs-on-Mac/bin:$PATH"
export FVP_MOUNT_DIR="$PWD"
FVP_Corstone_SSE-320 --version

    

Install released PyTorch and run Arm setup from the ExecuTorch repository root. The setup options select the cross-compiler and FVPs; the next section installs the Python export tools:

    

        
        
python -m pip install "torch==2.14.0" --index-url https://download.pytorch.org/whl/cpu
./examples/arm/setup.sh --i-agree-to-the-contained-eula \
  --disable-ethos-u-deps --enable-baremetal-toolchain --enable-fvps

    

Install ExecuTorch

Install the released ExecuTorch package and its Ethos-U dependencies:

    

        
        
python -m pip install "executorch[ethos_u]==1.5.1" "torchao==0.18.0" "cmake<4" pytest
source examples/arm/arm-scratch/setup_path.sh

    

The path script adds the downloaded cross-compiler and, on Linux, FVP binaries to the current shell. Check the target tools and Python installation:

    

        
        
command -v arm-none-eabi-g++
command -v FVP_Corstone_SSE-320
python -m pip check
python - <<'PY'
import executorch.codegen.tools.selective_build
from executorch.backends.arm.ethosu import EthosUPartitioner
from executorch.backends.arm.quantizer import EthosUQuantizer

print("ExecuTorch installation verified")
PY

    

The Python checks should report No broken requirements found. and ExecuTorch installation verified.

In each new shell, return to the ExecuTorch repository root and restore the environment and tool paths:

    

        
        
source .venv/bin/activate
source examples/arm/arm-scratch/setup_path.sh
if [ "$(uname -s)" = Darwin ]; then
  export PATH="$PWD/examples/arm/arm-scratch/FVPs-on-Mac/bin:$PATH"
  export FVP_MOUNT_DIR="$PWD"
fi

    

Download the model and sample audio

Create a directory for the model and audio inputs:

    

        
        
mkdir -p silero-vad

    

Download the model and sample audio from the tested Silero VAD revision:

    

        
        
curl --fail --location \
  --output silero-vad/silero_vad.jit \
  https://raw.githubusercontent.com/snakers4/silero-vad/dbacf536adadf42210f37ae50fbaf75f6235b3cf/src/silero_vad/data/silero_vad.jit

curl --fail --location \
  --output silero-vad/test.wav \
  https://raw.githubusercontent.com/snakers4/silero-vad/dbacf536adadf42210f37ae50fbaf75f6235b3cf/tests/data/test.wav

    

Create the validation clip

Use the example’s helper to take the first 40,000 samples of test.wav. This creates a 2.5-second validation clip that fits the runtime’s audio limit and produces 79 frames:

    

        
        
python examples/arm/silero_vad_example_ethos_u/model_export/generate_smoke_audio.py \
  --source-wav silero-vad/test.wav \
  --output silero-vad/validation.wav

    

Use test.wav for calibration and validation.wav for the host reference and FVP. These recordings overlap to check deployment consistency. Use representative calibration audio and separate validation recordings when measuring model quality.

What you’ve accomplished and what’s next

You’ve installed the pinned ExecuTorch source, prepared the Arm tools, and created the model inputs.

Next, you’ll export the model as a quantized ExecuTorch program for Ethos-U85.