Phonon proves speech synthesis can run locally in 65MB

Every voice interface stumbles on the same bottleneck: streaming audio to a remote server adds hundreds of milliseconds before the first syllable plays. Gradium's release of Phonon tackles that latency by fitting text-to-speech onto client hardware, compressing an autoregressive flow-matching architecture into a 65 megabyte footprint. The 31 million parameter English model and its 40 million parameter multilingual sibling generate audio latents locally across CPU, WebGPU, and mobile NPUs rather than waiting on cloud inference.

According to results on Hugging Face's Open TTS Leaderboard, that small footprint puts Phonon at the top of the CPU streaming benchmark for time-to-first audio while staying competitive on word error rates against models up to one hundred times larger. Flow matching on discrete audio latents has usually been prone to numerical instability during training, especially when forced into tiny parameter budgets. Managing to stabilize that across five languages without inflating memory requirements shows how much headroom remains in architectural efficiency over brute force scale.

For teams building interactive voice systems, moving audio synthesis to the edge eliminates API costs and keeps sensitive speech data on the device. The trade-off sits in the licensing: while the English checkpoint is Apache 2.0, the multilingual weights remain under a non-commercial license. That split will test whether developers adopt a bespoke runtime for an open English voice or wait for unconstrained weights in French, German, and Spanish.