vall-e/vall_e.cpp

vall_e.cpp

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

At the moment it's very barebones as I try and wrestle with llama.cpp's API without needing to modify its code.

Build

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

Populate ./libs/ with the compiled libraries of llama.cpp and encodec.cpp.

• encodec.cpp requires updating ggml to the latest version and doing a quick hack to make it work on the CPU backend.
• llama.cpp currently requires no hacks, but:
  • would be very nice to retrieve a model's tok_embd through the API.
  • would be very nice to only specify a slice of the output head through the API.

Run make.

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
  • the tokenizer is being a huge thorn on actual sequences
• load audio from disk
• encode audio
• sum embeddings for the prom and prior resps
• AR sampling
• NAR-len demasking sampling
• NAR sampling
• decode audio to disk
• a functional CLI
• actually make it work
  • it seems naively stitching the model together isn't good enough since the output is wrong, it most likely needs training with a glued together classifier