ChatGPT
Perplexity
Claude
Gemini
Grok
What Is ZONOS3-Mini and How Does Browser Voice Cloning Work?
ZONOS3-Mini is Zyphra’s text-to-speech model for reading a script in a reference voice. For a Pexo video project, think of the result as a narration track. You supply Zyphra’s model with the words you want spoken and a sample of the speaker. Its browser edition can generate audio on the device itself through WebGPU. Whether the first line plays quickly depends on the hardware, browser, and whether the model is already loaded. Zyphra’s release announcement
What does “Mini” mean here?
Zyphra released ZONOS3-Mini on November 9, 2026. Its sparse mixture-of-experts design has 800 million parameters in total, with 100 million active during an inference step. The distinction matters if you are estimating a download or deciding whether to test a phone. Only part of the model participates in each computation, while the runtime still needs access to the weights it uses across the whole generation. The active parameter count alone cannot tell you the download size. Zyphra model card
| Release detail | What it tells a creator |
|---|---|
| Zero-shot speech generation | Try a new voice using a reference recording without training a dedicated model for that speaker. |
| WebGPU runtime | A compatible browser can perform the speech calculation locally. |
| Public weights on Hugging Face | Developers can choose a deployment route and inspect the distributed artifacts. |
| Apache 2.0 model license | The model card identifies the terms for the weights. Permission to use a person's recorded voice is a separate question. |
How does ZONOS3-Mini clone a voice?
Imagine recording yourself saying “The studio opens at nine” and then typing “Our afternoon session begins at two.” The typed sentence supplies the content of the new recording. The audio reference supplies information about how you sound. Generating an accurate sentence and generating a recognizable voice are separate achievements. Listen first for whether the words came out correctly, then compare the speaker's resonance, vowels, and cadence with the reference.
The native generation API accepts a reference clip or the identifier of a saved speaker embedding. An embedding is a numerical representation that lets the model condition later generations on the speaker. A saved voice is convenient for a series of narration pickups. For example, keep the same voice selection while regenerating a single line whose price or date changed in the script. Official generation API and examples
| Input or setting | What to keep fixed when testing | What a change helps isolate |
|---|---|---|
| Script | The same short sentence | Pronunciation problems tied to a particular word |
| Reference recording | One clean clip of a speaker who gave permission | How much the choice of clip affects voice similarity |
| Delivery | A neutral reading | The effect of a requested emphasis or whisper |
| Playback | The same headphones and volume | Whether a difference survives a level-matched comparison |
Can ZONOS3-Mini run in a browser?
Yes. Zyphra provides an in-browser build that uses WebGPU. Open its local playground on the device you plan to use and allow time for the initial load. Choose a supplied voice first so you can test generation before recording anything. Type a sentence short enough to remember and listen to the first word and the last one. Then repeat the same sentence. That second run gives you a more relevant sense of making revisions once the session is ready. Open the official browser playground
Several deployments share the Voice Studio interface. A page displaying a text box and a play button therefore tells you little about where inference happens. The official repository distinguishes the browser build, a local server on Apple Silicon, and a server running on an NVIDIA GPU. If your requirement is on-device speech, start from the link specifically labeled “Run in your browser.” Zyphra deployment guide
| Available route | Where the speech computation runs | A reason to choose it |
|---|---|---|
| Browser WebGPU build | Compatible client device | Audition a line on the target laptop or phone. |
| MLX runtime | Apple Silicon Mac | Set up a repeatable local workflow on a Mac. |
| Mini-SGLang backend | Supported Linux machine with NVIDIA GPU | Serve generation requests from an application. |
| Zyphra Cloud | Hosted infrastructure | Try the hosted model independently of local GPU support. |
What do the reported latency numbers measure?
Zyphra reports browser time to first audio below 570 milliseconds on a Pixel 9 and below 150 milliseconds on an M3 MacBook Air. These are the publisher’s results on named devices. The metric stops when audio first becomes available. A listener assessing a conversation will also care about whether playback continues smoothly. A creator exporting a long narration will care about when the complete file is ready. Zyphra’s device measurements
| Clock | Start and stop | Example use |
|---|---|---|
| Initial setup | Open the playground until the model is ready | Estimate the wait for a first-time visitor. |
| First audio | Submit a prepared request until speech starts | Assess responsiveness for a short interaction. |
| Full generation | Submit until the whole passage is available | Plan how long a batch of narration pickups takes. |
| Playback continuity | Listen from first syllable to last | Catch pauses that a fast start can hide. |
For a practical device test, use one short line followed by a paragraph from the intended script. Record which device and browser you used and whether you had already generated speech in the session. Time the first audible word separately from the end of generation. Then listen to the middle of the longer passage. A line that starts promptly can still contain an odd pause or a misread name that needs another take.
How can you hear the difference between identity and delivery?
Try saying “I said Tuesday” with the stress first on “I” and then on “Tuesday.” The first version corrects who spoke. The second corrects the date. A convincing clone should retain the speaker’s recognizable qualities across both readings. Zyphra’s release includes controlled listening examples where text and narrator stay fixed while one requested characteristic changes. The examples are useful when a setting such as accent or voice texture sounds abstract on a specification sheet. Listen to Zyphra’s control examples
For a narrated video, review the actual line against its scene. A plausible voice can still emphasize the wrong number or pause midway through a product name. Write down the exact word that needs attention, keep a copy of the first take, and compare the replacement at the intended playback level. Save the accepted line with its script revision so the editor can tell which reading belongs under the picture.
Related reading
- Compare AI voice-cloning tools
- Turn a finished audio track into video
- Explore an audio-to-video skill for Claude Code
Frequently Asked Questions
What is ZONOS3-Mini used for?
ZONOS3-Mini generates speech from text in a reference voice. A Pexo creator planning a narrated video could use an approved voice take as the starting audio for a separate production workflow.
Does it need a training session for each new narrator?
Zyphra describes the model as zero-shot. Supply a reference recording or select an existing voice, then generate the new line. No dedicated model-training run for that narrator is part of this workflow.
Can I use it without installing a Python server?
Yes. Zyphra provides a WebGPU browser playground. The self-hosted Python and native server options are separate deployment choices.
What is the difference between a voice sample and a script?
The voice sample gives the model a speaker reference. The script supplies the words for the output. To change a sentence while preserving the speaker selection, edit the text and generate another take.
Which phone was used for the published browser figure?
Zyphra names Google's Pixel 9 and reports less than 570 milliseconds to first audio. Test your own device before using that figure to schedule an interactive experience.
Does first-audio latency include the time to finish a whole paragraph?
First-audio latency ends as soon as the first audio becomes available. Time the complete passage separately if your application must wait for a finished recording.
Can I reuse a voice across requests?
Yes. The native API accepts a saved speaker embedding identifier. Keep the selected voice constant when comparing two versions of the same script.
Can ZONOS3-Mini run locally on Apple Silicon?
Yes. Zyphra documents an MLX server for Apple Silicon Macs. The browser edition is another route on a compatible device if you want to audition speech before setting up a server.
What should I listen for in a cloned voice?
Check the intended words, then compare the identity and delivery with the original speaker. A correct transcript can coexist with an unnatural stress pattern, so listen to the sentence in context.
Where are the weights and runtime code?
Zyphra hosts the model weights on Hugging Face and maintains the inference runtime on GitHub. Consult the relevant repository for the terms governing those particular artifacts.
How should I choose between browser and hosted generation?
Use the browser build when your first question is whether your own device can generate the line acceptably. Use the hosted playground to audition the voice without depending on that device’s local inference performance.
