TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #81

Turn long passages of text into natural, repeatable speech

Most speech models race toward realism with ever bigger checkpoints and GPU bills. This release, built and funded by one independent developer, asks the opposite question: how small can a complete, natural-sounding voice get? The answer fits in less space than a photo album.

MODELInflect-Micro-v2
PUBLISHEDJuly 27, 2026
READ TIME3 min
TESTED BYNeural Expedition
CATEGORYAUDIO

Field notes

01What it does

Inflect Micro v2 turns written text into spoken audio, start to finish, on a plain CPU. The whole model is 37 megabytes, and it still synthesizes about six times faster than real time, so a minute of narration takes about ten seconds on an ordinary machine. You get one fixed male English voice; there is no voice cloning and no reference audio to prepare, which is exactly what makes it a two-minute setup.

Two behaviors set it apart from other tiny speech models. It is deterministic: fix a seed number and the same sentence renders as the same audio every single time, so narration in an app or video stays stable across rebuilds instead of drifting. And it handles long text properly, splitting at punctuation and rejoining with natural pauses, so you can feed it pages rather than sentences. In a blind community listening test it was preferred 66.2 percent of the time over the established compact options like Piper and KittenTTS.

02How to try it

Open the live playground on Hugging Face and paste a sentence that usually trips small voices up, numbers and names included: "The package weighs twelve point six kilograms and arrived on July twenty-first." You will hear the answer in a few seconds, no account needed.

Running it locally is a genuinely small job: download the model from Hugging Face, install the requirements, and three lines of Python write a wav file, CPU only. An ONNX version runs without PyTorch at all, small enough for edge devices and even the browser.

03Caveat

One fixed male English voice is the whole offer: no cloning, no other languages, and an experimental adaptation toolkit only if you bring your own data. Expect "clearly best tiny voice", not studio-grade realism; unusual names and odd phrasing can come out flat, and the developer is honest that unfamiliar text is the failure mode to test first.

04What you can do with it

  • Add spoken alerts or status messages to an app with no cloud speech bill.
  • Generate voiceover for tutorials and internal videos entirely on your laptop.
  • Bake identical, repeatable narration into every build of a game or product demo.
  • Read long documents aloud offline, with nothing leaving your machine.
  • Put a voice on small devices where multi-gigabyte speech models cannot fit.

Try the demo

View model page

Read this issue on neuralexpedition.com →