Inflect v2 brings ultra-tiny TTS models under 4M and 10M parameters

A high-school developer has released Inflect v2, a pair of extremely compact text-to-speech (TTS) models that push the boundary of how much useful speech synthesis can fit into a tiny parameter budget. The two models—Inflect-Nano-v2 at 3.96 million parameters (15.97 MB FP32) and Inflect-Micro-v2 at 9.36 million parameters (37.53 MB FP32)—are complete end-to-end systems. Text goes in, 24 kHz speech comes out, with no external vocoder, hosted API, or separate learned components. Both run locally on CPU or CUDA via a PyTorch API. The Nano prioritizes minimal footprint, while the Micro uses extra capacity for better clarity and stability.

In blind community testing against other compact TTS systems, Micro and Nano finished second and third among tested voices. Micro achieved a 4.395 UTMOS22 score with 3.99% semantic word error rate (WER) and runs 6.28× real-time on CPU; Nano scored 4.386 UTMOS22 with 4.21% WER and runs 10.72× real-time. For context, Nano is roughly 21× smaller than Kokoro, 126× smaller than Chatterbox, and over 1,000× smaller than Fish Audio S2 Pro. Limitations include English-only, a single fixed male voice, no voice cloning, and occasional metallic artifacts. The developer emphasizes this is a size-to-quality tradeoff that's now genuinely compelling for edge deployment.