Building from source

TorchAudio integrates PyTorch for numerical computation and third party libraries for multimedia I/O. It requires the following tools to build from source.

• PyTorch

• CMake

• Ninja

• C++ complier with C++ 17 support

  • GCC (Linux)

  • Clang (macOS)

  • MSVC 2019 or newer (Windows)

• pkg-config (Linux/macOS, if building sox extension)

• CUDA toolkit and cuDNN (if building CUDA extension)

Most of the tools are available in Conda, so we recommend using conda.

• Building on Linux and macOS
• Building on Windows
• Building on Jetson

Customizing the build

TorchAudio’s integration with third party libraries can be enabled/disabled via environment variables.

They can be enabled by passing 1 and disabled by 0.

• BUILD_SOX: Enable/disable I/O features based on libsox.

• BUILD_KALDI: Enable/disable feature extraction based on Kaldi.

• BUILD_RNNT: Enable/disable custom RNN-T loss function.

• BUILD_CTC_DECODER: Enable/disable CTC decoder based on Flashlight Text.

• USE_FFMPEG: Enable/disable I/O features based on FFmpeg libraries.

• USE_ROCM: Enable/disable AMD ROCm support.

• USE_CUDA: Enable/disable CUDA support.

For the latest configurations and their default values, please check the source code. https://github.com/pytorch/audio/blob/main/tools/setup_helpers/extension.py