Hacker News
Inflect-Micro-v2: Apache-2.0 Text-to-Speech in 9.36M Parameters, 6.28× Real-Time on 4 CPU Threads
owensong's Inflect-Micro-v2 landed on HN at 180 points with only 15 comments — a VITS-family English TTS model with 9,356,513 deployable parameters and a 37.53MB FP32 footprint producing 24kHz mono audio. Reported quality: 66.2% preference in community blind listening tests, 4.395 UTMOS22 naturalness, 3.99% semantic error rate across multiple ASR evaluators, running 6.28× real-time on four CPU threads. Apache-2.0 licensed with 153 likes and 298 monthly downloads, though the training corpus pipeline and full optimization recipe are withheld, so it is a usable artifact rather than a reproducible one.
↳ Follow the thread