8perezm/esuyo-qwen3-tts-rocm

This repository packages Qwen3-TTS, a text-to-speech AI model, into a server you can run on your own computer, built specifically for AMD GPUs using ROCm instead of the more common NVIDIA CUDA setup. Text-to-speech means turning written words into spoken audio. The server copies the same API format that OpenAI uses for its own speech service, so anything already built to talk to OpenAI's text-to-speech endpoint can point at this server instead, with no code changes needed. Getting started requires Docker, a tool that packages software into containers so it runs the same way on any machine, plus an AMD GPU with ROCm support and at least 4 gigabytes of video memory. You start the whole thing with one command, docker compose up, which builds the container and launches the server, made available locally on port 8000. Once running, you send text to the server and get back an audio file. The README shows three ways to do this: a plain curl command from the terminal, a Python script using the requests library, or the official OpenAI Python SDK pointed at the local server address. Alongside the text you want spoken, you describe the voice you want in plain language, for example a deep, resonant male voice with a British accent, and the model tries to match that description rather than picking from a fixed list of named voices. The first time you start the server, it downloads the underlying 1.7 billion parameter model from Hugging Face, which can take several minutes, but afterward the weights are cached in a Docker volume so later restarts are quick. The README also explains a startup warning about a missing speed optimization called flash-attn, noting this is expected and not a problem, since the code falls back to PyTorch's built in attention handling, which runs nearly as fast without needing a difficult manual build step. This project has zero stars and no listed license, so treat it as an early, unproven, personal setup rather than a widely used or supported tool.