Gradium Open-Sources Phonon TTS Model Under Apache 2.0

The Phonon on-device TTS model is available in 31 million and 40 million parameter sizes. The English Phonon 31m model, released under the Apache 2.0 license, is contained within a 64.5 MB file and allows developers to generate natural speech without requiring cloud calls, GPUs, or incurring per-character fees.

The Phonon 31m model is English-only and includes 37 built-in voices, distributed under Apache 2.0. The Phonon 40m models come in two variants: one model per language (English, French, German, Spanish, Portuguese) under CC BY-NC 4.0, and a multilingual model supporting all five languages with 102 voices, also under CC BY-NC 4.0.

Phonon utilizes a causal transformer backbone with a prefix combining text and speaker embeddings. It employs a SentencePiece tokenizer with a 4,000-token vocabulary for raw text input, bypassing phoneme conversion. The backbone generates latent vectors every 80 ms, which are then processed by a flow-matching head to produce an audio embedding in a single step.

A variant of Kyutai’s Mimi decoder then generates 24 kHz PCM audio, with a separate end-of-speech head determining the output's termination.

The Phonon 31m architecture consists of 6 transformer layers with a width of 512 and 8 attention heads, while the Phonon 40m uses 9 layers at the same width. Both models were distilled from a larger Phonon teacher model. The training process was conducted on an NVIDIA H100 cluster, leveraging NCCL and CUDA Graphs, and builds upon Kyutai's prior work.