Is Phonon state of the art?Pending benchmarks.Announced Oct 20, 2026; no independent results yet.

The details so far

Lab

Gradium

Announced

Oct 20, 2026

API

Not yet

Open weights

Published

Parameters

31M and 40M (excluding the audio tokenizer)

Lab

Gradium

Models

Phonon 31m (English), Phonon 40m (one model each for English, French, German, Spanish and Portuguese) and Phonon 40m multilingual (all five)

Weights

Hugging Face: phonon-31m-en, phonon-40m-en/-fr/-de/-es/-pt, phonon-40m-multilingual, plus phonon-voices; q8 files 64.5 MB (31m) and 74.6 MB (40m)

License

Phonon 31m Apache 2.0; all 40m models CC BY-NC 4.0 (commercial use through Gradium sales); phonon-voices Apache 2.0

Builds on

Kyutai's Pocket TTS and the Mimi codec (model card)

Runs on

CPU, Apple Neural Engine (Core ML), CUDA, Metal, Vulkan, Snapdragon NPU, and the browser via WebAssembly / WebGPU (lab statement)

Hugging Face repos created

Oct 19, 2026 (UTC); announced Oct 20, 2026

Tags marked unsure are not stated on the lab's own pages.

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

• Publicly announced with a source
• Usable by the public (API or weights)
• Listed on an independent board we track
• Enough results for a rank in a category
• Holds across two snapshots (crowns only)

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: 1.70% word error rate for Phonon 31m on 500 hard English sentences (phone numbers, email addresses, postal codes), against 3.50% for Kokoro, 7.40% for Kitten TTS and 8.71% for NeuTTS Nano.

Sources