Gradium launches open-source on-device tts model with apache 2.0 license

Gradium, a Paris-based voice AI company, has released Phonon, an open-source text-to-speech (TTS) model under the Apache 2.0 license. The model, available in 31M and 40M parameter versions, runs locally on devices without requiring cloud connectivity, GPUs, or per-character billing. Phonon uses a causal transformer architecture and generates speech at 24 kHz PCM, with a 64.5 MB file size for the 31M variant. The model outperforms existing TTS systems in both accuracy and speed. On a public hard-sentence dataset, Phonon 31M achieved a 1.70% WER, surpassing Kokoro’s 3.50% and NeuTTS Nano’s 8.71%. On the Seed-TTS set, it scored 0.51% LLM WER, compared to Kokoro’s 0.45%. Phonon 31M also runs 29.9x faster than real time on a MacBook Air M5 CPU and 46.7x real time with Apple Accelerate. Gradium’s Phonon is built on Kyutai’s research and integrates with tools like Transformers, MLX-Audio, and Ollama. The company highlights its ability to run on diverse hardware, including Raspberry Pi 5 and Snapdragon 8 Elite Gen 5, with fast startup times and efficient performance.