Inflect-Micro-v2: a full local text-to-speech model in under 10M parameters

Independent developer Owen Song has released Inflect-Micro-v2, a self-funded English text-to-speech model that packs a complete text-to-waveform pipeline into 9.36 million parameters (about 37.5 MB in FP32) and outputs 24 kHz mono audio. It’s a VITS-family architecture with a phoneme frontend, monotonic alignment, stochastic latent synthesis, and an integrated neural vocoder, and it runs faster than real time on CPU as well as on CUDA. The model ships as part of a two-size family: Micro optimizes for quality below 10M parameters, while a Nano variant targets a sub-4M footprint, both behind the same API. Determinism via fixed seeds and punctuation-aware chunking for long passages are built in.

The release is notable for its unusually transparent evaluation. Rather than collapsing performance into one number, it reports several metrics separately: a 66.2% win rate in an anonymous blind-preference study (framed explicitly as informal community evidence, not formal MOS), a UTMOS22 predicted-naturalness score of 4.395, and word error rates around 2.5% on Qwen3-ASR and 5.45% on Nemotron across held-out prompts. Comparisons against compact baselines like KittenTTS Nano, Piper Low, and Supertonic 3 are hedged carefully — weight footprints, ASR discrepancies (Whisper’s hallucinations are disclosed rather than hidden), and runtime caveats are all documented, and the frozen protocol includes corpus hashes and bootstrap intervals.

The practical significance is that credible, natural-sounding TTS no longer requires large models or cloud services — this fits on modest hardware and runs entirely offline, with a separately published ONNX export for CPU/CUDA/DirectML and browser deployment without a PyTorch dependency. The weights are open, though Owen keeps the training-corpus construction and optimization details private, and he’s floated a broader v3 with more languages and voices if the project finds an audience.