Inflect-Micro-v2: Full TTS in 9.36M Parameters
Original: Inflect-Micro-v2: complete voice in 9.36M parameters
Why This Matters
Sub-10M parameter TTS capable of real-time CPU inference expands viable edge and on-device voice AI deployment options.
Developer Owen Song released Inflect-Micro-v2, a text-to-speech model achieving complete voice synthesis in just 9,356,513 parameters (37.53 MB FP32). The model runs on CPU or CUDA, outputs 24 kHz mono audio, and recorded a 66.2% human blind preference rate against established compact TTS baselines.
Owen Song independently built and funded Inflect-Micro-v2, a fixed-voice English TTS model that delivers full text-to-waveform synthesis in under 10 million parameters. The model weighs 37.53 MB at FP32 precision and achieves 6.28× real-time throughput on a 4-thread CPU, with 24 kHz mono output. It supports deterministic seeds and long-text handling, and is available under the Apache-2.0 license.
Inflect v2 ships in two sizes sharing a single public API: Micro targets quality below 10M parameters, while Nano prioritizes footprint below 4M. Evaluation used multiple metrics rather than a single score: UTMOS22 predicted naturalness reached 4.395, two-ASR semantic word error rate was 3.99%, and a blind community preference study of 34 trials yielded a 66.2% win rate (21 wins, 10 losses, 3 ties). Competitors benchmarked include KittenTTS Nano, Piper Low, and Supertonic 3. Song noted that if the release gains traction, a broader v3 with additional languages, voices, and stability improvements is planned.