Kyutai's Pocket TTS clones voice from 5 seconds on CPU, MIT license

Kyutai's Pocket TTS is a ~100M parameter streaming language model that generates audio tokens via the Mimi neural codec and decodes to 24kHz. Unlike traditional acoustic-model-plus-vocoder pipelines, it operates as an autoregressive LLM for audio, producing tokens one by one. This architecture yields a remarkably flat latency: its real-time factor (RTF) stays between 0.69 and 0.76 whether the input is 12 characters or 1,712 characters, with no fixed overhead. The headline feature is zero-shot voice cloning from just a 5-second reference clip on CPU—capturing accent, timbre, pacing, and even microphone character without fine-tuning. Benchmark results on an Intel Xeon 8272CL (4 cores, 16GB RAM) show Pocket TTS at RTF 0.714 and UTMOS 4.10, behind Kokoro 82M (RTF 0.665, MOS 4.46) and Supertonic 3 at 5 steps (RTF 0.240, MOS 4.32), but those models cannot clone voices at all.

Pocket TTS is the only CPU-friendly model offering user-supplied voice cloning, making it a category of one for applications requiring personalization on edge devices. In contrast, Kokoro and Inflect-Nano ship fixed voice sets, and Supertonic offers none. Among the six configurations tested, Supertonic 3 (2-step) is fastest at RTF 0.121 but with poor quality (MOS 1.53), while Inflect-Nano-v1 has a 15-second output cap and buzzy quality (UTMOS 3.48). For fixed-voice, highest-quality, the Kokoro 82M (PyTorch or ONNX) is the pick. But for voice cloning on CPU, Pocket TTS stands alone. Installation is trivial: 'pip install pocket-tts'—no CUDA, no complex wiring. Licensed under MIT, it's a practical breakthrough for developers building interactive, personalized TTS on cost-effective hardware.