vall-e/vall_e.cpp

vall_e.cpp

This is an implementation that makes use of llama.cpp and encodec.cpp.

Model weights can:

• be found at ecker/vall-e@gguf
• converted with vall_e.export --yaml=./model_path/config.yaml --hf, then running python3 /path/to/your/llama.cpp/convert_hf_to_gguf ./model_path/hf/

Build

Populate ./include/ with the ggml, llama.cpp, and encodec.cpp headers.

Populate ./lib/ with the compiled libraries of llama.cpp, encodec.cpp, and espeak-ng (if not already in your LD_LIBRARY_PATH).

Run make.

Required Modifications

encodec.cpp requires updating its GGML copy to the latest version, which requires a few lines to get the CPU backend working (per my fork).

llama.cpp only possible modification needs to ensure that a non-causal attention mask is used; everything necessary can be hacked together with clever tricks.

To-Do

• converted model to GGUF
  • convert it without modifying any of the existing code, as the tokenizer requires some care
• basic framework
  • load the quantized model
  • orchestrate the required embeddings
  • juggle the output head / classifier properly
• phonemize text
  • with the help of espeak-ng
• tokenize phonemes
  • tokenize with llama_tokenize instead of a homebrewed method because the tokenizer is being a huge thorn
• load audio from disk
• encode audio
• sum embeddings for the prom and prior resps
• working AR output
  • AR sampling
• working NAR-len output
  • NAR-len sampling
• working NAR output
  • NAR sampling
• decode audio to disk
• a functional CLI
• actually make it work
• clean up to make the code usable elsewhere
• configured to allow for being used as a lib
  • (I do need to validate this in my engine project, but that's in MSYS2)
• feature parity with the PyTorch version
  • vocos
  • additional tasks
    • stt
    • ns / sr
    • samplers