Inflect v3 released: Apache 2.0 local TTS family with a 34M-parameter main model
b111ue · reddit · 2026-11-12
Developer Owen Song released Inflect v3, a family of local text-to-speech models under Apache 2.0.
- Mini (34.31M params): the quality-focused flagship, with six voices, reference-voice tuning from recordings, 48 kHz enhancement, and vocal-event tags like [laugh], [sigh], [gasp]; runs on CPU and CUDA
- Micro (11.79M params): a smaller neural speech model
- Atom (157K params): an extremely tiny learned synthesis core
Parameter counts refer to synthesis cores only; frontends and optional components add more. A Hugging Face playground, model collection, and benchmarks are available; the author solicits reports of mispronounced sentences with repro settings.
Related event: Inflect v3 Released: Open-Source Local TTS Starting at 34M Parameters(3 posts)→
More from Models
- Contested claims that Claude wasn't trained on Pokémon: burden of proof cuts the other way — sebkrier · 2026-11-12
- OpenAI falls to 3rd on Epoch Capabilities Index as benchmark growth accelerates 1.9x — teortaxesTex · 2026-11-12
- In defense of slow Gemini: async coding favors strong models over fast-but-dumb ones — Aizkmusic · 2026-11-12
- OpenAI deactivates paid multi-seat business account citing "Abusive Behavior" after leaked key warning — MarrMooMoo · 2026-11-12
- AI houseshare stats: same-lab models name each other 1.64 vs 1.18 times on average — liminal_bardo · 2026-11-12
- Black-hat SEO insiders: Google manual actions rising, ChatGPT citations tightening — daluoseo · 2026-11-12