High-performance inference of Meta's Encodec deep learning based audio codec model:

• Plain C/C++ implementation without dependencies using ggml

Here is a demo of running Encodec on a single M1 MacBook Pro:

• Support of 24Khz model
• Mixed F16 / F32 precision
• 4-bit and 8-bit quantization
• Metal support
• cuBLAS support

• The core tensor operations are implemented in C (ggml.h / ggml.c)
• The encoder-decoder architecture and the high-level C-style API are implemented in C++ (encodec.h / encodec.cpp)
• Basic usage is demonstrated in main.cpp.

Here are the steps for the bark model.

git clone --recurse-submodules https://github.com/PABannier/encodec.cpp.git
cd encodec.cpp

In order to build encodec.cpp you must use CMake:

mkdir build
cd build
cmake ..
cmake --build . --config Release

Offloading to GPU is possible with the Metal backend for MacOS. Performance are not improved but the power consumption and CPU activity is reduced.

cmake -DGGML_METAL=ON -DBUILD_SHARED_LIBS=Off ..
cmake --build . --config Release

The inference can be offloaded on a CUDA backend with cuBLAS.

cmake -DGGML_CUBLAS=ON -DBUILD_SHARED_LIBS=Off ..
cmake --build . --config Release