"Kyutai's Pocket TTS: Studio-Quality Voice Cloning With Zero GPU Required"

 parameter count, CPU-only operation, ~200ms latency, 6-language support, open-source/pip-installable nature are all independently verified.

"quietly dropped" undersells that this had real press coverage since January; "6x faster than real-time" and the exact "5 seconds of audio" cloning figure weren't precisely confirmed in what I found (some sources cite 5–10s or 20s depending on version); and the "bypasses the usual token-transformer bottlenecks" line is a simplified/dramatized description of the architecture rather than a precise technical claim