Support for OuteTTS 1.0

[edwko]

Since v1.0 has simplified processing, this implementation provides full feature support.

Changes and Features

• JSON Speaker Loading:
  • Added support for the new JSON speaker format, which includes an interface version.
  • OuteTTS 1.0 is supported using interface version 3.
• Text Chunking for Long Inputs:
  • Enables processing of very long input texts by splitting them.
  • Splitting respects minimum and maximum word boundaries (min = 10, max = 30).
  • Supports multilingual text.
  • Can be disabled via --tts-no-text-chunking (default: enabled).
• Text Preprocessing & Prompt:
  • While optional, a light cleanup and normalization step is included to improve output quality.
  • Added new required prompt handling for the v1.0
• Code Organization:
  • Implementation is located in: tts-outetts-v1.cpp.
  • A default speaker is added in a header file as JSON default_speaker.h.

TODO / Help Needed

• DAC (Descript Audio Codec) Integration:
  • The decoder layers from DAC need to be implemented: descript-audio-codec/dac/model/dac.py
  • Model used:
    weights_24khz_1.5kbps_v1.0.pth
  • DAC is supported by the transformers library and can be converted to safetensors, which might help implementation.
    Also, see this PR I submitted to fix a dependency issue in the conversion script for compatibility with newer PyTorch versions:
    transformers PR #36393
  • Requesting assistance from @ngxson and @ggerganov for implementing this part.

Example Commands

Default generation uses default speaker automatically and chunked text:

build/bin/llama-tts-outetts-v1 -m "path/to/model.gguf" -p "A very very long text"

Disables chunked text:

build/bin/llama-tts-outetts-v1 -m "path/to/model.gguf" -p "Hello, how are you doing?" --tts-no-text-chunking

With custom speaker file:

build/bin/llama-tts-outetts-v1 -m "path/to/model.gguf" -p "A very very long text" --tts-speaker-file "path/to/speaker.json"

[ngxson]

• The decoder layers from DAC need to be implemented

FYI, currently we're missing Snake1d which should be implemented via https://github.com/ggml-org/llama.cpp/pull/12487

[ggerganov]

Does DAC replace WavTokenizer?

[edwko]

Does DAC replace WavTokenizer?

Yes, since this model is multilingual, DAC is a better fit for reconstructing audio across languages.

[Horschig]

It would be really great if this would get merged. However I was wondering whether it'd also be possible to add mulitlingual support to llama-server?

[foldl]

FYI: OuteTTS 1.0 is supported by chatllm.cpp. You can find DAC & SNAC implementation there.