Gradium drops phonon-31m-en: a 64.5 MB local TTS that runs inside browser tabs

phonon-31m-en is an ultra-compact, on-device text-to-speech model built to speak English smoothly without touching the cloud.

Its 8-bit quantized GGUF weights weigh just 64.5 MB, and the core architecture contains only 31 million parameters. It requires zero GPU compute, streaming directly inside laptop CPUs, mobile devices, or a single browser tab.

What it does

The model converts English text into 24 kHz mono speech with true streaming playback. Audio chunks start playing almost instantly while the rest of the sentence is still being predicted in the background.

That makes it ideal for local screen readers, offline voice assistants, game NPCs, web reader extensions, or Home Assistant setups that cannot afford the latency or bandwidth cost of a round trip to a remote server.

Gradium also hosts an open catalog of roughly 100 pre-packaged voices, letting developers swap voice profiles easily without retraining anything.

Under the hood and benchmarks

To achieve quality inside a 64 MB envelope, Gradium combined Pocket TTS design with Kyutai's Mimi neural audio codec. A flow-matching transformer predicts audio latent representations in 80 ms frames, which the Mimi decoder then turns into clean audio.

In blind pairwise human evaluations featuring 468 raters and over 8,000 votes, phonon-31m-en scored an Elo of 2098, beating Kokoro's 2043 and NeuTTS's 2034.

On hard structured sentences packed with dates, codes, and currency units, its word error rate measured 1.5%, significantly cleaner than Kokoro's 3.5% and Magpie's 28%. On the standard Seed-TTS benchmark, it posted a 0.51% error rate.

Keep in mind its limits: it only handles English, and very long inputs get split sentence by sentence, which can sometimes break prosody across paragraph boundaries.

Both the weights and the voice assets are published on Hugging Face under the Apache 2.0 license, allowing full commercial use. The runtime is equally open, shipping with Python packages, browser-ready npm bundles, and C++ bindings.

Why it matters

For indie developers and product builders, phonon-31m-en eliminates the cost and privacy friction of voice generation. Instead of piping user text to paid third-party endpoints or forcing users to download multi-gigabyte models, you can drop a 64 MB bundle directly into an app or browser.

We think this is a smart pivot away from the frontier parameter arms race toward practical edge utility. It will not replace giant studio-grade narration models for audiobooks, but for real-time interactive agents where latency and zero operational cost matter most, it sets a brand new baseline.