Zyphra launches browser-based speech synthesis with under 570 ms latency on Pixel 9

When a voice assistant has already generated an answer but is still silent, users notice every pause. One way to shorten the wait is to synthesize speech on the device itself. Zyphra has released ZONOS3-Mini: according to the company's measurements, the first audio arrives in under 570 ms in the browser on Pixel 9, and in under 150 ms on a MacBook Air with M3. An educational app, for example, could read hints aloud directly on the phone. Developers should test whether this approach meets their interface's requirements. Zyphra announcement.

First, a correction to the original Telegram post: the claim that there are “nine languages, with no Russian” conflicts with the current documentation. The model card lists 43 languages, including Russian. However, a language's presence on the list does not establish how well the model pronounces surnames, abbreviations and numbers in your product. Model card.

How this could help your product

Imagine an app for learning English. The user answers a question, receives a correction and hears a pronunciation example. If speech synthesis runs locally, the team has the option to remove the call to an external service from this step. How much that speeds up the entire conversation also depends on the rest of the system: speech recognition, response generation and text preparation.

ZONOS3-Mini can clone a voice from a sample and change its characteristics: perceived age, accent and timbre. It offers control over emotion, pace and the delivery of individual words. In a game, this lets developers try different intonations for a character's line; in education, it can emphasize a word a learner got wrong. These are advertised capabilities whose quality needs to be assessed by listening. Feature description.

A practical test is fairly straightforward: take actual interface text and compare several speech outputs in a blind test. Ask colleagues to rate intelligibility, how natural the pauses sound and whether the intonation fits. Tricky examples are particularly useful: a long subscription plan name, a monetary amount with its currency, an address or a Russian sentence containing an English term.

What Mini means under the hood

The model has 800 million parameters, but only about 100 million are active at each processing step. It uses a mixture-of-experts (MoE) architecture, distributing computation across specialized blocks. This helps reduce the work required during generation, but the model still needs to be loaded and stored. The active parameter count alone does not tell you how much memory the application will use. Technical description.

There are several integration options. The repository includes a browser version using WebGPU, a server for NVIDIA, an MLX version for Apple Silicon and a C++/ggml implementation. The servers share an HTTP API, and audio streaming is supported. A team can start with a browser prototype and then compare it with a local server on the target hardware. Code and instructions.

Running it in the browser requires an initial model download, after which the files are cached. A product therefore needs two separate measurements: a new user's first visit and a subsequent app launch. Fast generation after loading does not guarantee an equally fast first launch. Browser instructions.

What to test before deployment

The published latency figures are Zyphra's own results on specific devices. Treat them as a reference point for your own testing. Check time to first audio, uninterrupted playback of long responses, memory usage and how the phone behaves after several minutes of operation. Separately, examine what happens when switching tabs and running other workloads at the same time.

For Russian-language use, text preparation is particularly important. The documentation lists nine languages with dedicated normalization; all others are passed to the model as written. Russian is not among those nine. This means that converting dates, monetary amounts and abbreviations should be part of integration testing. Text processing description.

The weights are released under Apache 2.0. That makes the model a promising candidate for a custom voice interface, but deployment decisions are best made after building a small prototype with real utterances, target devices and a measurable threshold for acceptable latency. The central product question is whether users receive an intelligible spoken response quickly and consistently enough. Weights license.