Gradium open-sources Phonon, sub-50M on-device TTS models hitting 1.7% word error rate
mattturck · x · 2026-10-22
Gradium has open-sourced Phonon, a suite of on-device text-to-speech models all under 50M parameters, claiming new standards for naturalness and accuracy with the broadest inference support. Phonon-31m ships under Apache 2.0, runs locally on phones, laptops, embedded hardware and browsers, and needs as little as 64.5 MB.
Key numbers:
- On 500 hard sentences (phone numbers, emails, postal codes), Phonon-31m hits 1.70% WER in English, well ahead of Kokoro (3.50%), Kitten TTS (7.40%) and NeuTTS Nano (8.71%).
- The first game built on it, Space Jarv by the xybrid team, runs Phonon live inside a 3D game on a 16GB M1 Pro MacBook, voicing every line on-device with no cloud calls.
Gradium previously extended its seed funding to $100M with NVIDIA among new investors.
Related event: Gradium Open-Sources Phonon, Sub-50M On-Device TTS Models(16 posts)→
More from Multimodal
- Talking cat argues over breakfast and weaponizes mom's phone number in AI video — azed_ai · 2026-10-22
- Neon-Noir Duotone Portrait With Full Midjourney Prompt and Params — michaelrabone · 2026-10-22
- Midjourney artist shares elephant-headed traveller linework with SREF style code — tisch_eins · 2026-10-22
- Seedance 2.5 demo shows impressively realistic balance and gravity physics — CurieuxExplorer · 2026-10-22
- A 4-Minute AI Horror Short Film Made With Dreamina — AnduinHammerson · 2026-10-22
- AI-generated "date night with a creepy clown" video leans hard into the uncanny valley — Tz-bz · 2026-10-22