Phonon 31M (English)

Phonon 31M (English) is an open-weight text-to-speech model developed by Gradium. Designed for local, on-device inference, the model is optimized to run efficiently on standard CPUs without requiring a dedicated graphics card or cloud connection. It features a compact footprint, with quantized files measuring roughly 64.5 MB, and supports rapid generation speeds with minimal latency for real-time applications.

Architecture and Capabilities

The model utilizes a causal transformer backbone combining text and speaker embeddings, paired with a SentencePiece tokenizer that processes raw text directly without requiring explicit phoneme conversion. It generates latent vectors that are converted into 24 kHz mono audio frames using a flow-matching head and a compact decoder. Phonon 31M incorporates 6 transformer layers with a width of 512 and 8 attention heads, and it includes 37 built-in English voices under an Apache 2.0 license.