Gradium's Phonon: Open On-Device TTS Takes Off

Episode 1 · Gradium Open-Sources Phonon, On-Device TTS Models Under 50M Parameters (2026-10-20, 16 posts)

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 — all under 50M — and embeddable into any app with fully offline operation and no voice data leaving the device. The company claims a new on-device standard for multilingual naturalness and accuracy, along with the broadest inference backend support. The news was spread by @alexverem, @HeyAmit, @bigaiguy and others, while @pcuenq and @mishig25 noted the team is from France and has drawn Hugging Face attention.

Confirmed

• Phonon ships in two sizes, 31M and 40M parameters, all under 50M; models and code are open-sourced, with Phonon-31m under Apache 2.0.
• Accuracy: Phonon 31M (English) achieves WER of just 1.70%, tested on 500 hard sentences where other models often fail, excelling at phone numbers, emails and postal codes.
• Engineering details (per @lmoroney): the English 31M model is only 64MB, ships with 37 voices, and reportedly has 7ms first-audio latency on a MacBook Air.
• The inference stack is fully open source, written mostly in Rust, targeting 10+ backends including CPU, GPU and NPU; @VraserX cited voicearenaai benchmarks on Snapdragon platforms.
• Because it runs entirely locally, speech synthesis works smoothly even offline (tested by @heyamit), protecting privacy; it targets embedding into any app at zero usage cost, with sharers suggesting it can replace paid TTS services. @lmoroney also recommends it for generating local voiceovers when scripting or rough-cutting short videos.

Why it matters

• Sub-50M parameter sizes mean the models can be embedded into any app on resource-constrained devices like phones, without cloud dependency.
• Offline operation inherently avoids privacy risks from uploading voice data, suiting privacy-sensitive, high-volume synthesis use cases.
• A fully open-source Rust inference stack supporting 10+ backends lowers the barrier for deploying on-device TTS across hardware platforms.
• Low WER on structured text like phone numbers and emails addresses a long-standing pain point of traditional TTS.

• Gradium open-sources Phonon: sub-50M on-device TTS that keeps talking offline — HeyAmit_ · 2026-10-20
• Gradium open-sources Phonon, 31M-40M on-device TTS models with 1.7% WER — thisguyknowsai · 2026-10-20
• Gradium Open-Sources Phonon: 31M-Parameter On-Device TTS With 1.7% Word Error Rate — thisdudelikesAI · 2026-10-20
• Gradium open-sources Phonon on-device TTS: 31M model hits 1.7% WER — bigaiguy · 2026-10-20
• Gradium open-sources Phonon: 31M/40M edge TTS models with Rust inference stack for 10+ backends — l-m-z · 2026-10-21
• Gradium open-sources Phonon: sub-50M on-device TTS models under Apache 2.0 — thisdudelikesAI · 2026-10-21
• Gradium open-sources Phonon on-device TTS: 31m model hits 1.7% WER, beating Kokoro — alex_verem · 2026-10-21
• Gradium open-sources Phonon, sub-50M TTS models under Apache 2.0 — mishig25 · 2026-10-21
• Gradium open-sources Phonon, sub-50M on-device TTS models under Apache 2.0 — dee_hw · 2026-10-21
• Phonon: free local TTS with zero usage cost and full on-device privacy — thisdudelikesAI · 2026-10-21
• Gradium open-sources Phonon TTS models under 50M params with day-0 Transformers support — pcuenq · 2026-10-21
• Gradium Open-Sources Phonon: 31M On-Device TTS at 64MB, 7ms Latency — lmoroney · 2026-10-21
• Gradium open-sources Phonon: 64MB on-device TTS with ~7ms first audio, 37 voices — lmoroney · 2026-10-22
• Gradium open-sources Phonon TTS: 31M-param model is 64MB and speaks in 7.5ms on-device — VraserX · 2026-10-22
• Gradium open-sources Phonon TTS models under 50M params, smallest at 31M — lmoroney · 2026-10-22
• Gradium open-sources Phonon, sub-50M on-device TTS models hitting 1.7% word error rate — mattturck · 2026-10-22

Episode 2 · Gradium open-sources Phonon, an on-device TTS model for iPhone and Mac (2026-10-27, 3 posts)

Gradium open-sourced Phonon, a 31M-parameter on-device TTS model that runs fully offline on iPhone and Mac with 25 English voices, achieving a 1.7% word error rate and outperforming Kokoro in benchmarks.

• Gradium open-sources on-device TTS model Phonon with 1.7% WER, beating Kokoro — amos_gyamfi · 2026-10-27
• Gradium's open-weight TTS runs offline on iOS/macOS, benchmarked vs Kokoro and PocketTTS — amos_gyamfi · 2026-10-27
• Open-source 31M-param Phonon TTS runs 25 natural English voices locally on iPhone and Mac — amos_gyamfi · 2026-10-28

Episode 3 · Open-source on-device TTS Phonon tops 85K downloads in a week (2026-10-28, 2 posts)

Gradium's open-source on-device TTS Phonon has surpassed 85,000 downloads within a week. Its inference runtime is written in Rust, runs without PyTorch, and targets phones, laptops, browsers and local servers.

• Phonon on-device TTS hits 80K downloads with PyTorch-free Rust runtime — mattturck · 2026-10-28
• Open-source TTS model Phonon hits ~85k downloads in a week, Rust inference stack — mattturck · 2026-10-29