Kokoro TTS maker returns with 9 more distilled voices running 100x realtime on CPU
mesmerlord · reddit · 2026-10-29
- The Kokoro TTS developer is back with 9 new tiny voice models that run 100x realtime on CPU.
- Distilling Kokoro's other voices directly came out robotic — only default afheart held up — so he pivoted to distilling voices from Phonom (40M params vs Kokoro's 80M; full model is 2-4x realtime, 100MB fp32).
- Surprise finding: the agent ended up reusing basically the same phenomizer as Kokoro, trained on Phonom-rendered outputs.
- Long-term goal: clone a voice with a big model (BreezeTTS, Fishaudio), distill it into a phone-realtime mini model for his AI roleplay app — no per-character TTS fees, just a few hours of 4090 training per voice.
- Wren and Hugh sound best so far; others still slightly robotic but far better than PocketTTS-class models.
More from Multimodal
- Runway Dev launches on Codex for autonomous API integration — runwayml · 2026-10-29
- Few-Shotting a Procedural Lydian Mode Generator from Exported Ableton MIDI — teropa · 2026-10-29
- Atlas 3DGS delivers artifact-free table scans with multi-angle frames — theworldlabs · 2026-10-29
- NUS Proposes Wiener-Informed Decoding to Speed Up Pixel-Space Flow Model Training — NationalUniversityofSingapore · 2026-10-29
- NVIDIA open-sources Nemotron-Labs-Diffusion-Uni, unified diffusion LM for multimodal — siyan_zhao · 2026-10-29
- Seedance 2.5 on Pippit AI used to produce a cinematic brand promo, lowering the bar for video ads — anthara_ai · 2026-10-29