Gradium Open-Sources Phonon, an On-Device TTS Family With a 31M English Model Under Apache 2.0

The release covers three models. Phonon 31m is English-only, carries 37 built-in voices and ships under Apache 2.0, which permits commercial use. Phonon 40m is published as one model per language for English, French, German, Spanish and Portuguese, and a separate multilingual 40m model packs all five languages into a single network with 102 voices; both 40m variants are licensed CC BY-NC 4.0, which bars commercial use. The 40m files are 74.6 MB. Parameter counts for the family exclude the audio tokenizer, and output is 24 kHz mono audio.

According to the report, the models run on CPUs, phones, Raspberry Pi 5 boards, browsers through WASM and WebGPU, and NVIDIA GPUs, and do not require a GPU. Developers get speech generation without a cloud call or a per-character bill.

The performance figures published with the release come from Gradium's own evaluation. The company reports about 7 ms to first audio on a CPU and says Phonon holds the top English Elo score among the open models it tested. On a 500-sentence hard test set it reports a best word error rate of 1.70%, against 3.50% for Kokoro in the same evaluation. On the Seed-TTS LLM test set, Gradium reports a worst-case 0.51% word error rate, behind Kokoro at 0.45% and Magpie at 0.46%. MarkTechPost did not report independent verification of these numbers.

Architecturally, Phonon uses a causal transformer backbone whose prefix combines text and speaker embeddings. A SentencePiece tokenizer with a 4,000-token vocabulary reads raw text without phoneme conversion, and the backbone emits one latent vector per 80 ms frame. A flow-matching head converts that vector plus noise into an audio embedding in a single step, and a reduced version of Kyutai's Mimi decoder produces the 24 kHz PCM output. A separate end-of-speech head decides when generation stops. Phonon 31m uses six transformer layers at width 512 with eight attention heads; Phonon 40m uses nine layers at the same width. Both were distilled from a larger Phonon teacher model.

Training ran on an NVIDIA H100 cluster using NCCL and CUDA Graphs. The work builds on Kyutai's Continuous Audio Language Models research and on Pocket TTS.

Editor's Summary

Gradium has released the Phonon family of on-device text-to-speech models, with the English 31M version available under the permissive Apache 2.0 licence and the 40M multilingual versions restricted to non-commercial use. The models run without a GPU on phones, browsers and single-board computers, which removes cloud inference from speech generation. Gradium's reported word error rates and latency are its own evaluations, and the release positions Phonon against open competitors such as Kokoro.