Gradium open-sources the ultra-small Phonon speech-generation model; 31M English-version weights are only 64MB

Article

Gradium open-sources an ultra-small speech model: weights only 64MB, enabling speech even when a phone is offline

Beating Insights AI breaking news,AI speech company Gradium has open-sourced its speech-generation model, Phonon. Earlier this year, it April began private testing with approximately 100000000 parameters; this release introduces 31000000 and 40000000 parameter versions and publicly releases the model weights for the first time. Among them, the 31000000-parameter English version's Q8 weights are only 64MB, and it can run on an ordinary CPU. 40000000 The parameter version supports languages—English, French, German, Spanish, and Portuguese—and is available as both single-language and combined -language models.

Phonon is trained by distilling a larger model. It reads text directly without first converting it into phonemes, then uses Transformer to progressively predict acoustic features and generates audio through the Mimi decoder. 80 Each output corresponds to

milliseconds of speech and can be played while being synthesized, without waiting for the entire sentence to be generated. 8 Elite Gen 5 In terms of speed, Gradium disclosed that Phonon requires only 7.5 milliseconds to generate the first audio segment on Snapdragon 24, achieving a synthesis speed M5 times faster than real time. On the CPU of a MacBook Air 7, time to first audio is about

milliseconds. 1189 For audio quality, Gradium recruited 17000 listeners, who completed more than 31000000 blind-listening votes. 2098 Elo The parameter English version scored 100000000, higher than the TTS achieved by the parameter Pocket 2056. In another English test of 500 sentences containing complex text such as phone numbers and email addresses, Phonon's word error rate was 1.7%, lower than the 82000000 of Kokoro, another open-source speech model with 3.5%parameters. These results all came from tests published by Gradium.

Phonon has been integrated with tools including Ollama and MLX-Audio. Partner Xybrid also ran 16GB M1 Agent, Whisper, and Phonon simultaneously on a 8B Pro MacBook, enabling 3D game characters to converse in real time while completely offline.

Currently, only the 31000000-parameter English version uses the Apache 2.0 license, which permits commercial use. 40000000 The parameter and multilingual versions use the CC BY-NC license; commercial use requires separate authorization. No Chinese model has yet been released.

Original link https://m.theblockbeats.info/flash/372962

Key points

Gradium has open-sourced the Phonon speech-generation model and made its model weights public.
This release of Phonon introduces 31000000 and 40000000 parameter versions.
31000000 The Q8 weights of the parameter English version are 64MB, and an ordinary CPU can run it.
40000000 The parameter version supports English, French, German, Spanish, and Portuguese, and is available as both single-language and combined -language models.
Phonon reads text directly and uses Transformer to progressively predict acoustic features, after which the Mimi decoder generates audio.
Each model output corresponds to 80 milliseconds of speech and can be played while being synthesized.
Gradium states that Phonon requires 8 Elite Gen 5 milliseconds to generate the first audio segment on Snapdragon 7.5, achieving a synthesis speed 24 times faster than real time; on the M5 of a MacBook Air CPU, time to first audio is approximately 7 milliseconds.
Gradium recruited 1189 listeners to complete more than 17000 blind-listening votes; the 31000000-parameter English version scored 2098 Elo，Pocket TTS scored 2056。
In an English test of 500 sentences containing complex text such as phone numbers and email addresses, Phonon's word error rate was 1.7%, versus 3.5%。
for Kokoro. Phonon has been integrated with tools including Ollama and MLX-Audio.
Xybrid simultaneously ran 16GB M1 Agent, Whisper, and Phonon on a 8B Pro MacBook, enabling 3D game characters to converse in real time while completely offline.
Only the 31000000-parameter English version uses the Apache 2.0 license and permits commercial use; the 40000000-parameter and multilingual versions use the CC BY-NC license, and commercial use requires separate authorization. The article states that no Chinese model has yet been released.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Phonon is a small model that converts text into speech. Its English version has a small file size and can run on ordinary CPU while generating and playing audio simultaneously, making it suitable for local voice scenarios that do not require an internet connection. The language coverage and commercial-use licenses differ across versions.

Why it matters to readers

Smaller weights, the ability to run on ordinary CPU hardware, and streaming output can reduce local speech generation's requirements for devices and networks; however, language support and licensing must both be verified before adoption.

Beginners may wish to focus on the 31M English version: it is compact and permits commercial use, but there is currently no Chinese model. Speed, audio-quality, and accuracy data are primarily published by Gradium and should not be regarded directly as independently verified results.

Risks and unknowns

The speed, audio-quality, and word error rate results all come from tests published by Gradium; the original text does not provide independently reviewed results.
The original text does not provide an official model repository, weight download address, or complete reproduction method.
40000000 parameters and the multilingual version use the CC BY-NC license; commercial use requires separate authorization.
No Chinese model has been released yet.
The fully offline real-time conversation demonstrated by Xybrid pertains to specific hardware and configuration combinations; the original text does not describe performance on other devices.

Gradium open-sources the ultra-small Phonon speech-generation model; 31M English-version weights are only 64MB

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center