Kyutai has released Pocket TTS, a ~100M parameter model capable of cloning voices from just five seconds of audio. Benchmarked against Kokoro, Supertonic, and Inflect-Nano via UTMOS scores, the model is noted as the slowest of the tested configurations but remains highly significant in the CPU TTS field.