NeuTTS Air

NeuTTS Air is a small text-to-speech model from Neuphonic built for on-device deployment. It pairs a compact LLM backbone of roughly 360 million active parameters, about 550 million including embeddings, with NeuCodec, a 50 Hz single-codebook neural audio codec that keeps audio quality high at low bitrates. The model clones a voice from as little as three seconds of reference audio and generates natural-sounding speech in real time on mid-range consumer hardware, which makes it suitable for embedded voice agents, assistants, toys, and other applications that need local audio generation without calling cloud APIs. Input text is converted to phonemes before synthesis, and generated audio carries a watermark. It ships in PyTorch form plus GGUF quantizations at Q4 and Q8 for efficient local inference, and the model is released under the Apache 2.0 license.