Kokoro Web
🇺🇸 English

Updated · Practical guide

Kokoro TTS in Your Browser: Setup, Voice Selection, and Measured Results

Kokoro Web uses the Kokoro-82M model to turn English text into a WAV file in your browser. This guide explains the current workflow, what the site actually supports, and one measured run with enough detail to repeat it. The timing is a single-device observation, not a promise about every computer or a voice-quality ranking.

What Kokoro Web currently does

The site loads the Kokoro-82M v1.0 model into a browser worker. When WebGPU is available, the app selects its WebGPU runtime; otherwise it uses a WebAssembly path. The model files must be fetched before the first generation and may be cached in browser storage for later sessions. Cache behavior and startup time vary by browser, connection, and available storage.

Once the model is ready, the entered text is passed to the local worker for synthesis. This means the text does not need to be uploaded to a speech-generation server. It does not mean the entire website is network-free: the page, model files, and separate site services can still make requests. The current tool accepts up to 2,000 characters and offers 54 voices across English, Mandarin Chinese, Japanese, Spanish, French, Hindi, Italian, and Brazilian Portuguese. Choose a voice matching the text language and export generated speech as WAV. Japanese first use also downloads a pronunciation dictionary.

Generate and check a WAV file

  1. Open the Kokoro TTS tool and wait until its status says the model is ready. The status also identifies whether this session selected WebGPU or WASM.
  2. Paste up to 2,000 characters of English text. Use normal spelling and punctuation. The current text box does not interpret SSML tags or bracketed acting notes as controls.
  3. Choose a voice and a speed between 0.5× and 2.0×, then generate. Listen for names, dates, abbreviations, and sentence breaks; revise the text if a phrase is unclear.
  4. Download the WAV and make exact pauses, loudness adjustments, or edits in an audio editor when the project needs them.

For a long script, split at natural topic or sentence boundaries so each part stays within the input limit. There is no measured evidence here for a universal “best” segment duration, loudness target, or voice-to-content formula, so choose those settings by listening to the intended material.

One measured browser run

On October 29, 2026, I used this repository’s local development build in Chrome on macOS. The app reported WebGPU and “Model ready” before timing began, so the measurement starts after model initialization. I entered the 421-character sample below, kept the default Heart · en-us · Female voice and 1.0× speed, clicked Generate, and stopped the timer when the completed audio player appeared. I read the player’s audio duration after generation.

Before publishing a narration, read the text aloud once and listen for names, numbers, and sentence breaks. This sample uses ordinary English punctuation so the same input can be repeated on another device. The browser downloads the speech model when it is first needed, then runs this synthesis locally. Keep the original text, choose one voice and speed, and listen to the full WAV before editing the script for rhythm.

Observed result for this single run
MeasureObserved value
Runtime and settingsWebGPU · Heart · en-us · Female · 1.0×
Input length421 characters
Click to completed player6.33 seconds
Generated WAV duration27.78 seconds
Generation time ÷ audio duration0.23×

This single model-ready run shows that this setup produced about 27.8 seconds of audio in about 6.3 seconds. It excludes the first model download and page startup. The hardware model and exact Chrome build were not captured, and no WASM or cloud provider was tested with the same text. Treat this as a reproducible example, not a general speed claim or evidence that one service sounds better than another.

How to make a fair comparison

Text-to-speech tools differ in language, voice, speed controls, network path, and export format. A fair performance test holds the input, voice style, output settings, device, and network conditions steady; it reports both generation time and audio duration; and it repeats each run after clearly labeling cold and cached starts. A voice-quality comparison also needs a listening rubric and multiple listeners. The single test above does not meet that standard, so this guide does not claim that Kokoro is faster, more natural, or better than cloud services overall.

  • Privacy: this app synthesizes text in the browser after the model loads. Check another provider’s own documentation for its upload and retention practices.
  • Startup: Kokoro Web requires model files on first use; browser caching can reduce later loading, but is not guaranteed to persist.
  • Language and controls: the current generator offers 54 voices across eight languages and plain-text controls. Check another tool’s supported languages and markup before choosing it.
  • Voice quality: listen to the same representative script in the voices you would actually use. Parameter count alone does not establish which voice sounds better.

Limitations and related guidance

WebGPU availability depends on the browser, operating system, and graphics hardware. If the app selects WASM instead, expect a different runtime and do not reuse the WebGPU timing above as a prediction. Very long scripts may need several generations because of the 2,000-character input cap. The browser can also clear cached model files, so keep a network connection available when you need to reload them.

Use the Kokoro voice guide to review the available voices and the model author’s grading notes. For sound cleanup after export, see our audio quality guide. The privacy policy explains analytics, model-host, and advertising requests that are separate from speech synthesis.

Frequently asked questions

Does Kokoro Web work offline?

A first model download is required. Browser storage may retain the files, but it can be cleared and the site may still need to load other assets or services, so offline operation is not guaranteed.

Can I enter SSML, pauses, or emotion tags?

The tool accepts plain text. SSML and bracketed stage directions are not special commands. Use punctuation to guide phrasing and edit precise pauses after downloading.

Does this test compare WebGPU with WASM?

No. It records one model-ready WebGPU run. A valid comparison would repeat the same script and settings on both runtimes and report the device, browser, cache state, and several trials.

Is the entered text sent to a speech server?

The current app runs synthesis in a browser worker on your device. The page and model still need to load, and the site may make separate analytics or advertising requests; see the privacy policy for those services.