Inflect v3: open-source local TTS with a 157K-param model, Apache 2.0, 0.74% WER

pmttyji · reddit · 2026-11-17

owensong released Inflect v3, a family of compact Apache 2.0 English TTS models for local inference: Mini (34.31M, most features), Micro (11.79M), and Atom (157.19K, signal-processing-based synthesis). All offer six preset voices and 24kHz output; Mini adds reference-voice tuning, 8 event tags, and optional 48kHz enhancement. Self-run benchmarks for Mini: 0.739% average WER over 1,688 texts, UTMOSv2 3.917, 162x real-time on an RTX 5090. The author notes these are not official leaderboard results. Code and demos are on GitHub and HF Spaces.

Original post →

More from Multimodal

Multimodal channel →