Ternary quantization shrinks VoxCPM2 5.3x to 827MB, runs real-time on laptop CPU
oguzeldereli · reddit · 2026-11-18
Developer oguzeldereli converted the 2B open-source TTS model VoxCPM2 from bf16 (4.96GB) to ternary 1.58-bit weights (-1/0/+1) with a Q8 diffusion transformer, yielding an 827MB model (+100MB VAE) — 5.3x smaller — served by a custom Python-free C++ runtime.
Quality largely holds:
- Blind listening score 3.02, beating full-precision VoxCPM2's 2.75
- UTMOS 4.31 vs 4.25; English WER 2.61%; speaker similarity 0.956
- 5–20s voice cloning similarity 0.933, near the 0.936 of two real clips from the same speaker; cloning WER 1.9% matches full precision
- Zero-shot synthesis across 27 languages (FLEURS)
Speed: RTF 0.16 on RTX 4070 Ti SUPER; real-time on a laptop i9-14900HX CPU only; RTF 2.5 on older CPUs without VNNI. Model (Apache-2.0), runtime and conversion code are all open-sourced. The same TTF method was previously applied to LLM training/memory reduction.
More from Multimodal
- Real-time 3D worlds with Gaussian splats: 'the barriers of imagination are dissolving' — pvncher · 2026-11-18
- APOB AI pitches Seedream 5.5 workflow for AI idol album covers and K-pop MV teasers — aftahi_ai · 2026-11-18
- Looped layers don't help multimodal embeddings unless recurrence is pretrained (+4.23 points) — _reachsumit · 2026-11-18
- Building a Claude skill for anime-style emotional framing techniques — andrew_n_carr · 2026-11-18
- sprite32: open-weights 32x32 text-to-sprite diffusion model trained from scratch — DaneMarguglio · 2026-11-18
- Claude makes videos at near-zero token cost: user builds automated film production engine — RileyRalmuto · 2026-11-18