Kyutai's Pocket TTS Clones Voices in 5 Seconds, on CPU, Under MIT

Kyutai's recently released Pocket TTS operates as a roughly 100M-parameter streaming language model rather than the typical acoustic-model-plus-vocoder pipeline, generating audio tokens sequentially over Kyutai's Mimi neural codec before decoding to 24kHz. This architecture yields a flat real-time factor (RTF 0.69-0.76) regardless of text length, and its streaming nature makes it well-suited for interactive applications.

The model's standout capability is zero-shot voice cloning from just a 5-second reference clip, running entirely on CPU with no fine-tuning or GPU required—capturing accent, timbre, pacing, and even microphone character. Competing CPU-friendly models like Kokoro, Supertonic, and Inflect-Nano all ship with fixed voice sets, making Pocket TTS the only option in this category for CPU-based custom voice cloning.

In a benchmark spanning 180 timed runs, 36 audio samples, and objective UTMOS-based MOS scoring, Pocket TTS was the slowest of six tested configurations (RTF 0.714), yet its cloning ability sets it apart. Kokoro scored highest on quality (UTMOS ~4.44-4.46) but lacks cloning; Supertonic prioritizes speed but carries OpenRAIL-M licensing restrictions; Inflect-Nano is tiny but its UTMOS score is inflated by a known scoring quirk, actually sounding buzzy and robotic, plus it has a hard 15-second output cap.

A key detail for engineers is Pocket TTS's MIT license and frictionless pip installation with no CUDA build required. For developers seeking commercial flexibility, this is a meaningful advantage over Kokoro's Apache 2.0 and Supertonic's more restrictive OpenRAIL-M terms.