Kokoro gains voice cloning via OpenVoice with ~600MB extra VRAM

A developer added a Kokoro plus OpenVoice ToneColorConverter recipe to audio.cpp, giving existing TTS models voice cloning without retraining. Cost is about 600MB extra peak VRAM, roughly 10% slower CUDA and about half speed on CPU but still 2x real-time. CAM++ speaker similarity was 0.7-0.8, with results varying by built-in voice and reference audio quality.

• audio.cpp update cuts Higgs Audio TTS VRAM use by 48%2026-10-08