With Phonon, Gradium brings text-to-speech to everyday devices

Gradium releases Phonon, 31M to 40M parameter voice models capable of generating speech locally on desktop, mobile, and browser.

A synthetic voice capable of starting to speak in a few milliseconds, without sending text to a remote server. With Phonon, Gradium is releasing a family of speech synthesis models compact enough to run directly on a computer, smartphone, or in a browser. The smallest has 31 million parameters, while the 40-million variants cover five languages, including French. The company highlights a latency of 7.5 milliseconds on the Snapdragon 8 Elite Gen 5 and generation up to 24 times faster than real-time on this chip. Available with its weights on Hugging Face, Phonon seeks to make quality speech synthesis accessible without permanent cloud dependency, with applications ranging from video games to personal assistants and accessibility tools. Speech Synthesis That Stays on the Device

Most advanced speech synthesis applications rely on remote services: text is transmitted to a model hosted in a data center, and then the generated audio is returned to the application. This operating model allows the use of large models, but introduces network dependency, communication delays, and costs associated with API calls. For a video game character, an assistive interface, or a connected device, these constraints can become noticeable when responses must be produced frequently.

Phonon approaches the problem from the opposite direction. Its models are designed to be loaded and executed directly on the hardware used by the application. Phonon-31m has 31 million parameters, with an 8-bit quantized weights file of 64.5 MB. The 40m versions feature 40 million parameters and occupy approximately 74.6 MB in their quantized representation. These relatively contained sizes make it possible to target machines that do not necessarily have a dedicated GPU.

This approach builds on research conducted by Kyutai around Pocket TTS, which had already shown that good quality speech synthesis could run in real-time on a consumer CPU. Gradium adopts certain principles of this architecture and seeks to improve pronunciation accuracy, voice diversity, and support for multiple languages. The company emphasizes three properties: text remains on the device when inference is fully local, the service can continue to operate offline, and generation does not require payment for every use. Seven Milliseconds Before the First Sounds

The claimed performance relies primarily on the delay before the first audio segment and on the ratio between the duration of the produced sound and the time required to compute it. On a MacBook Air M5, Gradium measures approximately 6.9 milliseconds before the first sound for Phonon-31m using four CPU threads. In this configuration, the model generates the equivalent of 29.9 seconds of audio per second of computation. With additional hardware-specific optimizations, this ratio reaches 46.7 times real-time in the company's tests.

On Arm processors, the integration of KleidiAI, an optimized math library developed by Arm, improves performance. Arm reports a delay of 28 milliseconds before the first sound and generation fifteen times faster than real-time. In collaboration with Qualcomm, Gradium also announces a delay of 7.5 milliseconds on the Snapdragon 8 Elite Gen 5, with audio production 24 times faster than the playback of the result.

These measurements depend on the hardware, the number of threads, available optimizations, and how the model is executed. They do not represent the full response time of a conversational assistant: the eventual transcription of the question, the language model's reasoning, and the text preparation add to the voice computation. They nonetheless demonstrate that once the text is available, synthesis can begin with a low enough delay to no longer systematically require a round trip to a server. French, German, Spanish, and Portuguese Versions

The Phonon family includes several configurations. Phonon-31m-en, the most compact model released under the Apache 2.0 license, specializes in English. The Phonon-40m variants each have a version dedicated to English, French, German, Spanish, and Brazilian Portuguese. A 40-million-parameter multilingual variation also makes it possible to produce all five languages from the same weights, without needing to load a separate model for each language.

Gradium compared these models against several open systems, including Kokoro, Magpie, NeuTTS, and Pocket TTS, using blind listening tests. The naturalness evaluation relies on 1,189 participants and 17,270 collected preferences. In the published results, Phonon-40m-en achieves an Elo score of 2,117 in English, compared to 2,043 for Kokoro. In French, Phonon-40m-fr reaches 2,072 points, while the multilingual version scores slightly lower.

Pronunciation constitutes another important criterion, particularly when texts contain phone numbers, dates, zip codes, or numerical values. Gradium assembled a set of 500 challenging sentences per language, whose outputs are evaluated using speech recognition systems and an LLM-assisted comparison. Phonon-40m-fr shows an error rate of 2.4% on this French dataset, compared to 5.5% for Kokoro in the reported results.

These evaluations provide indications regarding perceived naturalness and reading accuracy, but do not cover every situation. The tests rely on selected voices and sentences, and certain results vary depending on the protocol used. They do not allow for the conclusion that a Phonon variant will systematically outperform all larger models. A Generator That Produces Audio in Small Sequences

Phonon's architecture relies on a causal Transformer combined with representations of the text and the chosen voice. The model processes textual instructions with a vocabulary of 4,000 tokens, without having to systematically convert sentences into phonemes. It then produces successive audio representations at a rate of one sequence every 80 milliseconds of sound.

A component based on flow matching transforms each representation into acoustic information, before a decoder derived