neonbjb/tortoise-tts

Tortoise TTS is a text-to-speech program. You give it some written text and it speaks the text out loud as an audio file. The author built it with two priorities in mind: handling many different voices well, and producing speech that sounds realistic in its rhythm and intonation. This repository holds all the code needed to run the system in inference mode, meaning you use the already-trained model rather than train your own. The name is a joke about speed. The README explains that the model is slow because it uses two stacked decoders, both of which sample audio at low rates. On older graphics hardware it could take about two minutes to generate a medium sentence. A later note in the README says speed has since improved, with a real-time factor of 0.25 to 0.3 on a 4 GB graphics card and latency under 500 milliseconds when using streaming. To use it locally you need an NVIDIA GPU. The README walks through a conda-based install with PyTorch, transformers, and the project itself. There is also a Docker recipe that drops you into a ready-to-use container, and separate instructions for Apple Silicon Macs using a nightly PyTorch build, with the caveat that one acceleration library called DeepSpeed does not work on those machines. Once installed, several command line scripts are provided. One speaks a single phrase, another reads long text files sentence by sentence and stitches the clips together, and a third runs a socket server on port 5000 for streaming use. The README also shows a small Python snippet for calling the model from your own code, with optional flags for half-precision math and key-value caching to run faster.