NederlandsSmall open AI model reads text aloud on your phone, without internet
The AI company Gradium has released Phonon as open models: small text-to-speech models that read text aloud with a natural voice, entirely on a phone, a laptop or in the browser. On a public Hugging Face leaderboard, Phonon is among the most intelligible text-to-speech models.
Computers read text aloud more and more naturally. But the best voices usually run on large servers in a data centre. Your text then travels back and forth over the internet. That takes time and money, and your data leaves your device. The AI company Gradium shows that it can be done differently.
Gradium was founded by the founders of the research lab Kyutai and is backed by, among others, the chipmaker Nvidia. On 21 October it released Phonon: a family of small text-to-speech models that run entirely on the device itself, without an internet connection. Anyone can download the models. According to a blog post on the AI platform Hugging Face dated 5 November, they were downloaded more than 200,000 times in the first two weeks alone.
Small enough for a phone
An AI model is made up of adjustable numbers, called parameters. The fewer it has, the less memory and computing power it needs. The English-language Phonon 31m has 31 million, plus about 10 million more for the part that turns the model's output into audible sound. The well-known open text-to-speech model Kokoro has 82 million. The file for Phonon 31m is 64.5 megabytes.
There are two versions. Phonon 31m speaks English and anyone may use it for free, including in commercial products. The slightly larger Phonon 40m speaks English, French, German, Spanish and Portuguese. It is free for non-commercial use; companies pay an amount that depends on their number of users.
The model runs on phones, laptops, small single-board computers and in the browser, and also on Nvidia graphics cards. For this, Gradium wrote its own software that works with more than ten types of computing chips, from ordinary processors to the special AI chips in phones and laptops. According to Gradium, on a powerful graphics card the first bit of speech is heard after just 5 milliseconds, and on a server processor after 43 milliseconds.
Mistral earlier released Voxtral 3, another model that reads text aloud, in more than 35 languages. It works via Mistral's servers or on a company's own computers. Phonon makes the opposite choice: fewer languages, but small enough to fit on a phone.
How it works
Phonon builds on earlier work by Kyutai. The model produces speech in pieces of 80 milliseconds. For each piece it predicts a compact description of the sound, and a second component turns that description into real audio. Because it works piece by piece, it can start talking before the whole text is ready.
Phonon has more than 100 built-in voices. With Gradium Studio you can also create your own voice: from a short recording of someone, only with that person's consent, or by describing in ordinary words how the voice should sound. The speaking speed can be adjusted. According to the makers, the model also reads phone numbers, dates and codes correctly.
How good is it?
Hugging Face keeps a public leaderboard for open text-to-speech models, the Open TTS Leaderboard. It compares models on intelligibility, speed and naturalness. Intelligibility is measured by automatically transcribing the spoken text again and counting the words that were misheard. On this measure Phonon ranks first in German and Spanish, and in the top three in every language tested, among models that are up to a hundred times larger. In the speed test on an ordinary processor, Phonon is the fastest of all models to start speaking.
Gradium also ran its own listening tests, in which people compared two voices at a time without knowing which model was behind them. For English, 468 listeners took part. According to Gradium, they chose Phonon more often than Kokoro and the text-to-speech models Magpie and Pocket TTS.
Anyone who wants to hear it for themselves can try Phonon in the browser.
What does this mean for the road to the singularity?
Anyone who talks to an AI assistant wants it to be fast and natural, and preferably without every word first travelling to a data centre. Phonon shows that a good voice fits into a small model that works on a phone or in the browser. That makes it easier to build reading aloud into apps, tools and devices, even where there is no internet and without texts leaving the device. The more of what an assistant does, such as listening and speaking, can happen on the device itself, the closer an AI helper comes that is always at hand, everywhere.
What is the technological singularity? Read the explanation →
