How we prepared voice cloning TTS to run on-device

The team at VoxRT has developed a compact, 45M-parameter text-to-speech (TTS) model capable of voice cloning entirely on-device. Designed to run on smartphones and Raspberry Pi, the model avoids cloud dependencies by utilizing a two-stage architecture: a language model for generating audio latents and a neural codec decoder for waveform synthesis. By replacing iterative diffusion with a single-step flow sampler, the developers achieved faster-than-realtime inference on budget hardware. The project highlights the engineering challenges of on-device AI, specifically addressing the tendency of single-step samplers to produce robotic, averaged sounds and the stability risks associated with mixed-precision training. The resulting system, which supports streaming and low-latency synthesis, is scheduled for release in September 2026. This development marks a significant step in making high-quality, personalized voice synthesis accessible without the need for server-side infrastructure or API calls.