Inflect-Micro-v2 delivers complete local text-to-waveform speech synthesis in 9,356,513 parameters (37.53 MB FP32), outputting 24 kHz mono audio. Built and funded independently by Owen, the model handles long text, deterministic seeds, and runs on CPU or CUDA. A companion Inflect-Nano-v2 (3.97M params, 15.97 MB) prioritizes minimal footprint.

In community blind preference testing, Inflect-Micro-v2 achieved a 66.2% win rate (21 wins, 10 losses, 3 ties) against baselines including Kitten TTS Nano, Piper Low, and Supertonic 3. It scores 4.395 UTMOS22 predicted naturalness and a mean 3.57% WER across Qwen3-ASR and Nemotron 3.5 on unseen text.

CPU throughput reaches 6.28× real-time on 4 threads (Hugging Face CPU Upgrade instance), while Nano hits 10.72×. Both models include an integrated waveform decoder and share the same Python API. Installation uses `hf download owensong/Inflect-Micro-v2`. Owen notes that likes on Hugging Face help discovery and may fund a broader v3 with more languages and voices.