Neutts-air – open-source, on device TTS (github.com)

Neuphonic released NeuTTS Air, an open-source, on-device text-to-speech system that combines real-time performance with instant voice cloning. Built around a compact 0.5B LLM backbone (Qwen 0.5B) and a proprietary neural audio codec (NeuCodec), NeuTTS Air is provided in GGML/GGUF formats and engineered to run on phones, laptops and even Raspberry Pis. The project emphasizes low-latency, low-power inference, producing highly realistic human-like voices and allowing users to clone a speaker from as little as 3 seconds of reference audio — unlocking privacy-friendly, offline voice agents, toys, and compliance-sensitive applications that don’t rely on cloud APIs. Technically, NeuTTS Air uses a simple LM + codec architecture tuned for the speed/size/quality tradeoff: a 0.5B text model for understanding/generation, NeuCodec with a single codebook for efficient low-bitrate audio, and optional ONNX or llama-cpp pipelines for CPU/GPU acceleration. The repo includes examples, encoder/decoder flows, and requirements (python≥3.11, espeak for phonemization), and recommends 3–15s mono WAV references at 16–44 kHz. Outputs are watermarked (Perth watermark) as a mitigant for misuse, but the instant cloning capability still raises important safety and policy considerations for the community as on-device voice synthesis becomes widely accessible.