ChatTTS tutorial: local setup with Ollama and Llama 3

This tutorial, uploaded on June 4, 2024, covers local installation of ChatTTS and an Ollama workflow identified in the source chapters as using Llama 3. The source gives no ChatTTS or Ollama version number, so the setup reflects that upload period.

The presenter clones the ChatTTS repository, installs its Python requirements and tries the repository's basic example. After an error, he adds imports for Torch and torchaudio and generates an audio file. He then tests advanced examples with laughter and break controls, shortens the input and increases the laughter setting to seven to hear how the output changes.

Repeated runs produce different voices and delivery styles. The presenter says he has not found a way to choose the voice; this is a limitation of his experiment, rather than proof that voice control is unavailable. He praises the speech quality but reports generation taking about 30 seconds despite using CUDA. He considers that setup too slow for real time speech interaction.

The final demonstration uses a local LLM through Ollama to prepare text for speech synthesis. A prompt asks the model to insert emotion tokens and includes examples of their use. The presenter plays several generated responses, though he says the intended token insertion does not appear to have worked. The video is useful as an early open source TTS experiment, with observed results rather than a performance benchmark.