[audio.cpp] Portable C++/GGML implementation of Higgs Audio v3 TTS: 8.8–10.1× real-time (warmed) and 8.5× real-time (longform).

Hi Higgs Audio team,

First, thank you for open-sourcing Higgs Audio TTS!

We recently finished an implementation of Higgs Audio TTS in audio.cpp and wanted to share the current performance results.

On an RTX 5090:

Warmed requests: approximately 8.8×–10.1× faster than real time Longform generation (6000+ chars): approximately 8.5× faster than real time

The implementation focuses on efficient local inference while maintaining output quality comparable to the reference implementation.

Check: https://github.com/0xShug0/audio.cpp