Gradium Open-Sources Phonon: Sub-50M On-Device TTS with 1.7% WER

Gradium has released and open-sourced Phonon, a family of text-to-speech (TTS) models designed to run on-device, available in 31M and 40M parameter sizes—both under 50M—so it can be embedded into any app with fully offline operation and no voice data ever leaving the device. The company claims it sets a new on-device standard for multilingual naturalness and accuracy, with the broadest inference backend support.

Confirmed

Phonon comes in two sizes: 31M and 40M parameters, all under 50M; the models and code are open source, with Phonon-31m released under Apache 2.0.
Accuracy: Phonon 31M (English) achieves a word error rate (WER) of just 1.70%, tested on 500 hard sentences where other models often fail, with standout performance on error-prone cases like phone numbers, emails, and postal codes.
The inference stack is fully open source, written primarily in Rust, targeting more than 10 backends including CPU, GPU, and NPU.
Because everything runs locally, speech synthesis works smoothly even offline, protecting privacy.

Why it matters

A sub-50M parameter count means the model can be embedded into any app on resource-constrained devices like phones, without relying on cloud services.
Offline operation inherently avoids the privacy risks of uploading voice data, making it well suited to privacy-sensitive embedded scenarios.
A fully open-source inference stack written in Rust and supporting 10+ backends lowers the barrier for developers deploying on-device TTS across different hardware platforms.

2026-10-20 ~ 2026-10-21 · 6 related posts

Primary sources

Gradium open-sources Phonon: 31M/40M edge TTS models with Rust inference stack for 10+ backends — l-m-z · 2026-10-21 00:01 (UTC+08:00)
Gradium open-sources Phonon, 31M-40M on-device TTS models with 1.7% WER — thisguyknowsai · 2026-10-20 23:33 (UTC+08:00)

Gradium open-sources Phonon: sub-50M on-device TTS that keeps talking offline — HeyAmit_ · 2026-10-20
[source] Gradium open-sources Phonon, 31M-40M on-device TTS models with 1.7% WER — thisguyknowsai · 2026-10-20
[source] Gradium open-sources Phonon: 31M/40M edge TTS models with Rust inference stack for 10+ backends — l-m-z · 2026-10-21

3 near-duplicate retellings: thisdudelikesAI · bigaiguy · thisdudelikesAI