Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models

Gradium has open-sourced Phonon, a family of compact on-device text-to-speech models designed to provide fast, natural speech without network access across CPUs, GPUs, NPUs, browsers, phones, and embedded devices. The 31-million-parameter English model, licensed under Apache 2.0, occupies 64 MB in q8 format and prioritizes a small footprint, while 40-million-parameter language-specific and multilingual variants support English, French, German, Spanish, and Portuguese with greater expressivity and accuracy under CC BY-NC licenses. Phonon includes text normalization for structured content such as dates, phone numbers, and version numbers, more than 100 built-in voices across variants, and compatibility with cloned or designed custom voices created through Gradium. The company reports that blind listener tests and structured-text word-error-rate benchmarks place its models ahead of comparable small TTS systems, while performance tests show low time to first audio and faster-than-real-time synthesis on hardware including MacBooks, Raspberry Pi devices, Snapdragon platforms, and web browsers. Built on a causal transformer, flow-matching audio generation, and a Mimi-based decoder, Phonon is distilled from a larger teacher model and can be extended through custom models for additional languages, specialized vocabularies, or constrained hardware applications such as automotive systems. Weights, inference code, and integrations for Python, Rust, Swift, JavaScript/WASM, MLX, Transformers, LM Studio, and Ollama are available through Hugging Face and GitHub.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.