# Inflect-Micro-v2 — 9.36M-parameter local text-to-speech under 40 MB

> A complete English text-to-speech model that fits in 37.5 MB and beats larger systems 66% of the time in blind human tests.

- Org: Owen Song
- Date: 2026-07-24
- Category: model, tool
- Source: https://huggingface.co/owensong/Inflect-Micro-v2
- Canonical page: https://ai-tldr.dev/releases/owensong-inflect-micro-v2/

## In short

Inflect-Micro-v2 is a 9.36M-parameter Apache-2.0 English TTS model that runs on a CPU at 6.28× real time, scoring 66.2% wins in blind human tests against larger systems.

## What it is

Inflect-Micro-v2 is a 9.36M-parameter English text-to-speech model from indie developer Owen Song. The full FP32 checkpoint is 37.53 MB, small enough to bundle inside a desktop or mobile app, yet it produces natural-sounding 24 kHz mono audio without needing a reference voice at runtime.

## How it works

The Inflect-Micro-v2 pipeline reads punctuation to segment long text, then generates the waveform directly on CPU or CUDA. Owen Song reports 6.28× real-time throughput on a 4-thread CPU, deterministic output from a seed value, and controls for playback speed (0.5–2.0×) and prosody variation, all through a single Python API and Hugging Face model download.

## Why it matters

Inflect-Micro-v2 shows that useful, natural TTS no longer needs a server: a 37 MB Apache-2.0 model that runs offline unlocks voice for accessibility tools, embedded devices, and any app whose creator does not want to route audio through a cloud vendor. In blind pairwise tests it won 21 of 34 comparisons, tying 3 and losing 10.

## Who it's for

indie app developers, accessibility tool builders, on-device voice tinkerers

## Try it

hf download owensong/Inflect-Micro-v2 --local-dir Inflect-Micro-v2

## Links

- [Model Card](https://huggingface.co/owensong/Inflect-Micro-v2)
- [Live Demo](https://huggingface.co/spaces/owensong/Inflect-v2)
- [GitHub](https://github.com/owenawsong/Inflect)
- [Hacker News](https://news.ycombinator.com/item?id=49053375)
