NeuTTS

NeuTTS is a collection of on-device text-to-speech speech language models from Neuphonic, distributed through a GitHub repository and a Python package. It synthesizes speech from text, optionally cloning a voice from a short reference audio clip, and runs locally using PyTorch or quantized GGUF backbones.

What It Is

NeuTTS combines small LLM backbones with the NeuCodec neural audio codec to generate speech locally rather than through a web API. The README lists model families including NeuTTS-Air (English), the NeuTTS-Nano multilingual collection (English, French, German, Spanish), and NeuTTS-2E, an English model with emotional control. Outputs are audio files written at 24 kHz, and every generated file carries a Perth watermark by default.

Voice Cloning and Emotion Control

Cloning requires a reference .wav sample and its transcript. The README recommends mono audio, 16-44 kHz, 3-15 seconds long, clean and continuous. NeuTTS-2E instead offers four fixed speakers (emily, paul, sophie, steven) and supports angry, disgusted, fearful, happy, neutral, sad and surprised emotions. Generation is sampled, and a seed can be set to reproduce identical audio.

Running On-Device

Install with pip, with optional extras for llama-cpp-python (GGUF models) and onnxruntime (ONNX codec decoder). Q4 and Q8 GGUF quantizations are provided, and streaming synthesis works with GGUF backbones. For lower latency, the README suggests GGUF backbones, pre-encoded references and the ONNX decoder. Example scripts cover basic synthesis, streaming, emotional TTS, and fine-tuning. The README publishes throughput benchmarks on CPU devices and an RTX 4090.

Licensing Notes

NeuTTS-Air is listed under Apache 2.0, while the Nano and 2E models use the NeuTTS Open License 1.0, which limits commercial use by entities at or above a stated revenue threshold. The README also warns that sites such as neutts.com are not affiliated with Neuphonic.