TEST RECORD · ← all field reports

FIELD REPORTS/AUDIO/ISSUE #130

Turn English text into natural speech that starts playing right away

Good text-to-speech usually means a cloud service or a big graphics card. This one is small enough that the demo downloads it into your browser tab and speaks from there, and in its maker's blind listening tests it still beat larger open voices.

MODELPhonon 31M
PUBLISHEDOctober 22, 2026
READ TIME3 min
CATEGORYAUDIO

Field notes

01 What it does

You type English text and it speaks it back in a clear, natural voice. Playback starts while the rest of the sentence is still being generated, so there is no wait before you hear the first words. It comes with about 100 voices to choose from.

What makes it different is the size. The whole model is a 65 MB download, small enough to run in a browser tab, on a phone or on a laptop processor, so your text never has to leave the device. It also handles the awkward parts of written English, reading numbers, dates, prices and codes the way a person would say them. In the developer's blind listening tests, people preferred it over Kokoro, a popular open voice model, and it made fewer mistakes on sentences packed with numbers.

It comes from Gradium and builds on Kyutai's Pocket TTS. This English version is released under Apache 2.0, so you can use it in commercial work.

02 How to try it

Open the Phonon Space. It runs entirely in your browser: the model downloads once, and nothing you type is sent anywhere. Choose Phonon 31M, then give it a hard line: "Your flight UA 482 leaves at 6:45 on March 3rd from gate B12, and the fare is $318.50." Listen for whether the flight number, time and price come out the way you would say them. Then paste a paragraph from an article you are reading and notice how quickly the first words arrive. To use it in your own project, the ptts Python package turns a line of text into a WAV file with one command, and there are ready guides for web apps, iPhone, Android and Home Assistant.

03 Caveat

It speaks English only, and text in other languages comes out mispronounced. Long passages are split into sentences and read one after another, so the tone does not carry across a whole paragraph. Sister models for French, German, Spanish and Portuguese exist, but those are licensed for non-commercial use only.

04 What you can do with it

  • Add a read-aloud button to a website or web app without paying for a voice service.
  • Give a game or interactive app spoken lines that play instantly, even offline.
  • Build a private voice assistant or Home Assistant setup where nothing goes to the cloud.
  • Turn articles or notes into audio you can listen to on a walk.
  • Power a screen reader or accessibility tool on a phone.

Try the demo

View model page

Read this issue on neuralexpedition.com →