wtf is higgs_audio_tts_easy_gui?

Higgs Audio TTS Easy GUI is a local desktop web interface for running Higgs Audio, a text to speech and voice cloning model, entirely on your own computer instead of through a cloud service. It wraps a full workflow: preparing reference voice clips, generating speech with either the newer Higgs V3 model or the older V2 model, transcribing audio with Faster-Whisper, building training datasets, and training custom LoRA voice adapters. The app moves a user through five stages in order. First, a person records or imports short reference voice clips, trims them, and transcribes them automatically or by hand, building a reusable library of saved voices. Second, the inference stage lets a user pick a model, optionally load a trained LoRA adapter, choose a reference voice, and generate speech from typed text, with longer text automatically split into chunks. Third, a dialogue builder assigns different voices to different speaker turns and stitches the results into one multi-speaker conversation, useful for podcast-style demos. Fourth, a dataset preparation tool converts a folder of audio clips into a structured training and evaluation dataset, filling in missing transcripts automatically. Fifth, a training interface runs LoRA fine-tuning on a chosen dataset. V2 training is treated as stable, while V3 training is marked experimental, since no official V3 fine-tuning method existed when this project was built. Installation uses a tool called uv to manage dependencies and comes with a Windows batch installer that sets up a virtual environment and installs the right version of PyTorch for the user's GPU, with several CUDA versions supported alongside a CPU-only mode and an experimental AMD path. After installing, launching the app opens a local web page in the browser. Running the heavier features needs a capable GPU. Faster-Whisper transcription can run on a small GPU or CPU, but Higgs V2 speech generation wants at least 12 gigabytes of video memory, Higgs V3 wants 16 gigabytes or more, and LoRA training wants 24 gigabytes. CPU-only mode works but is described as slow. The project also supports an optional PyTorch compilation mode that can more than double generation speed after an initial warmup pass. It is written in Python and aimed at people experimenting with local voice cloning and speech synthesis.