Beating AI News - AI voice company Gradium has open-sourced its speech generation model Phonon. It began private testing in April this year with approximately 100 million parameters, and now releases versions with 31 million and 40 million parameters, publicly releasing model weights for the first time. The 31-million-parameter English version with Q8 quantization weighs only 64MB, enabling operation on ordinary CPUs. The 40-million-parameter version supports five languages—English, French, German, Spanish, and Portuguese—and offers both single-language and multilingual models.
Phonon is trained via distillation from larger models. It directly reads text without first converting it into phonemes, then uses a Transformer to gradually predict audio features, which are finally generated as audio through the Mimi decoder. Each output corresponds to 80 milliseconds of speech, allowing real-time playback during synthesis without waiting for the entire sentence to be generated.
In terms of speed, Gradium reports that Phonon generates the first audio segment in just 7.5 milliseconds on Snapdragon 8 Elite Gen 5, achieving a synthesis speed 24 times faster than real-time. On MacBook Air M5's CPU, the first audio takes about 7 milliseconds.
Regarding audio quality, Gradium conducted blind listening tests with 1,189 participants, resulting in over 17,000 votes. The 31-million-parameter English version scored 2,098 Elo, surpassing the 100-million-parameter Pocket TTS at 2,056. In another test involving 500 complex English sentences including phone numbers and email addresses, Phonon achieved a word error rate of 1.7%, lower than the 3.5% of another open-source model, Kokoro, with 82 million parameters. All results are based on Gradium's published data.
Phonon has already been integrated into tools such as Ollama and MLX-Audio. Partner Xybrid also ran an 8B Agent, Whisper, and Phonon simultaneously on a 16GB M1 Pro MacBook, enabling real-time dialogue for a 3D game character even under complete offline conditions.
Currently, only the 31-million-parameter English version is licensed under Apache 2.0, permitting commercial use. The 40-million-parameter and multilingual versions are licensed under CC BY-NC, requiring separate authorization for commercial use. No Chinese language model has been released yet.