Paris-based voice AI startup Gradium, a spin-out from Kyutai, has released its Phonon text-to-speech model family under the Apache 2.0 license. The models are designed for local execution on devices including smartphones, laptops, and browsers.
The Phonon architecture utilizes a causal transformer backbone with a SentencePiece tokenizer featuring a 4,000-token vocabulary. The system generates audio at 24 kHz using a flow-matching head and a decoder derived from Kyutai’s Mimi technology.
The 31M parameter version features 6 transformer layers, while the 40M version uses 9 layers. Both models were distilled from a larger teacher model and trained on an NVIDIA H100 cluster using NCCL and CUDA Graphs.
In performance testing, the 31M model achieved a 1.70% word error rate on a public test set, compared to 3.50% for Kokoro and 8.71% for NeuTTS Nano. On a MacBook Air M5, the model runs 29.9x faster than real-time.
The software supports Rust, Python, Swift, C, and C++ environments, with compatibility for Transformers, MLX-Audio, Ollama, and LM Studio. It also provides an OpenAI-compatible server and a WebSocket server for integration.
