Encodec.cpp brings Meta's EnCodec audio codec to C++ with no dependencies
A lightweight C++ implementation that compiles weights into the binary for single-thread performance.
A Reddit developer has released Encodec.cpp, a clean C++ port of Meta's EnCodec neural audio codec backed by the Eigen linear algebra library. The project aims to provide a lightweight alternative to heavyweight ML runtimes like PyTorch or ONNX Runtime. Key design choices include compiling all model weights directly into the binary, eliminating runtime weight files, and ensuring zero external dependencies beyond a C++17 compiler and Eigen. This makes integration into CMake-based projects straightforward and enables maximum single-thread performance.
Functionally, Encodec.cpp supports the full EnCodec pipeline: high-quality audio compression at various bitrates and a learned audio tokenizer suitable for downstream tasks like speech synthesis or music generation. It handles dynamic input sizes (though no batching) and shows performance comparable or superior to onnxruntime in the developer's tests. The project is currently open for feedback, aiming to refine features like memory usage and potential multi-threading support. For developers needing efficient, dependency-free audio codec capabilities in C++, this offers a compelling option.
- Zero runtime dependencies: only requires C++17 and Eigen; no ML framework needed.
- Weights compiled into binary: no separate model files to manage or distribute.
- Performance matches or beats onnxruntime on single-thread, with support for dynamic audio sizes.
Why It Matters
Enables efficient, dependency-free audio compression and tokenization in C++ without heavy ML frameworks.