wtf is gepard-train?

This repository holds the training code for Gepard, a text-to-speech model that turns written text into spoken audio. Rather than being a tool you install and speak to right away, it is the pipeline researchers use to actually build the underlying model from raw audio data, in three stages: an initial pretraining pass over a large speech dataset, a fine-tuning pass that adapts it to short phrases, and an optional final stage that improves the naturalness of its output using preference comparisons between generated samples. Gepard's underlying design goal is to generate speech in a single fast pass rather than the multiple passes many older speech models need, so it can run at real-time or faster speeds on standard AI serving software. It also supports cloning a voice from a short reference audio clip. A separate, related repository handles running the finished model to actually produce speech, this repository is only for producing that model in the first place. Using it requires a computer with an NVIDIA GPU and Python, since it depends on CUDA for the heavy computation involved in training. A setup script installs the required packages into a local environment. From there, a set of simple commands builds a tokenized training dataset from audio, runs the pretraining stage across multiple GPUs, and runs the shorter fine-tuning stage on a single GPU. All the configuration for each stage lives in a structured settings system, so different training experiments can be run by swapping in different configuration presets. Once a checkpoint is trained, other commands merge it into a final servable model and upload it to Hugging Face for others to use. The project is released under the Apache 2.0 license. The full README is longer than what was shown.