neuphonic/neutts

NeuTTS is a collection of open-source text-to-speech models that convert written text into spoken audio. The key design goal is that the models run entirely on a local device, without sending data to a cloud service. The project is made by Neuphonic, a company focused on on-device voice AI. There are two model families. NeuTTS-Air is the larger one, with around 360 million active parameters, and supports English. NeuTTS-Nano is smaller at around 120 million active parameters and comes in separate versions for English, French, German, and Spanish. Both families support voice cloning: given just a few seconds of a person's audio, the model can generate new speech that sounds like that person. All generated audio is watermarked. The models are distributed in GGUF format, which is a file format commonly used for running AI models on ordinary computers without specialized hardware. Quantized versions are available at different quality levels to trade off file size against audio quality. The README includes benchmark results showing generation speeds on a Samsung Galaxy phone, a laptop CPU, an Apple M4 chip, and an NVIDIA graphics card. On a mid-range laptop the smaller model produces audio faster than real time. The Python package is installed via pip and can work with the llama-cpp-python inference backend for GGUF models. The README covers installation steps for different operating systems including macOS, Linux, and Windows, with notes on enabling hardware acceleration on each platform. Fine-tuning scripts are also included in the repository for those who want to adapt the models to custom voices or use cases. NeuTTS-Air is licensed under Apache 2.0. The NeuTTS-Nano models use a separate license called the NeuTTS Open License 1.0.