Higgs TTS 3 local setup and voice cloning tests

This tutorial sets up Higgs TTS 3, a 4-billion-parameter speech model, with an 8-bit GGUF version through audio.cpp. The presenter chooses Q8 because running BF16 causes problems while recording the screen. They review build instructions for Linux, Windows and Apple M-series Macs, then demonstrate the Mac Metal build and a Gradio interface for local AI speech tests.

Voice cloning uses reference audio and its transcript. Qwen3-ASR-0.6B supplies the transcript when that field is blank; entering one skips transcription. The walkthrough explains how Python calls the audio.cpp CLI and passes the model path and synthesis arguments.

The presenter finds the cloned voices similar to the references and tests English reference audio with French text, then French reference audio with English text. They praise the results but say they do not speak French and cannot confirm the French transcription's accuracy. The model card claims support for 102 languages, with 85 described as polished and production quality; these tests do not establish that wider coverage.

Expressiveness is the main disappointment in this setup. The presenter reports that sound-effect and style tokens fail to produce the intended delivery, even after adding the suggested onomatopoeia. Whispering and singing tests also disappoint. These observations concern the demonstrated Q8 workflow rather than every configuration of the model.