offline-speech

Download once, bookmark this page, and turn text into speech with no internet. Your text and audio stay on this device.

Next step

Preparing offline storage.

1. Download speech files

Start here: press Download for offline use. This saves the model and voices to disk once; it does not load them into RAM.

Checking saved files…

Offline files are stored in this browser profile. Browser settings or storage pressure can clear them.

2. Load the model

After the download finishes, press Load model. Wait for “Ready”, then go to step 3. Load it again whenever you reopen this page or unload RAM.

Disk · saved offline files
Checking…
RAM · model and runtime
Model unloaded
The page still uses some memory. Loading speech typically needs about 1 GiB of RAM on a desktop; this is an estimate, not a device reading.
Unloading keeps your saved files for quick offline reuse. Deleting removes the downloaded model, voices and libraries and unloads RAM. Only the small page shell remains; offline speech will need another download.
Storage details

Saved-file sizes exclude browser bookkeeping and temporary files.

3. Type text and hear speech

Choose a voice and press Speak. Speak prepares and plays the audio. Recent text-and-voice combinations replay instantly from the beginning without loading or running the model. Save WAV keeps the completed file. Progress stays below while it is preparing.

Save WAV

Waiting to speak.

Word counts advance after each sentence or long-sentence chunk completes. The model does not report individual-word completion.
Sentence breaks and progress
    Current phonemes

    Preparing offline storage…

    Licenses and dependencies

    Collated from Kokoro-82M, kokoro-js, Transformers.js, phonemizer/eSpeak NG, ONNX Runtime, Hugging Face Jinja, and PRN UI.

    offline-speech comes with no warranty; you may redistribute it under GPLv3 or later. offline-speech’s own code is licensed under GPL-3.0-or-later (GNU General Public License version 3 or later); third-party components retain their respective licenses. View the GPL license text.

    eSpeak NG: Copyright (C) 2005 to 2014 by Jonathan Duddington; Copyright (C) 2015-2017 Reece H. Dunn. Additional upstream copyright and license notices are provided in the full notices.

    Browser code, model and voice dependencies
    ComponentVersion / revisionLicense
    PRN UI8a11eb419d3c12f699c312068f622378d3307b50MIT
    Kokoro-82M model and six voices1939ad2a8e416c0acfeecc08a694d14ef25f2231Apache-2.0
    kokoro-js1.2.1Apache-2.0
    @huggingface/transformers3.8.1Apache-2.0
    phonemizer JavaScript wrapper (adapted)1.2.1 (6835144b7ee9043129222549c1ed2f6a27216278)Apache-2.0
    eSpeak NG code and language data (built by offline-speech)890071ff30231883d761e4d59edabd6d69faf159GPL-3.0-or-later
    Emscripten runtime7.0.0 (777e6e5db7b4990bcaeb1b9998ed642cc84e6629)MIT OR NCSA
    musl libc and LLVM libc++/libc++abi/compiler-rtVendored by Emscripten 7.0.0musl: MIT and component notices; LLVM: Apache-2.0 WITH LLVM-exception
    Unicode character data in eSpeak NG ucd-toolseSpeak NG 890071ff30231883d761e4d59edabd6d69faf159Unicode data/software license (COPYING.UCD)
    ONNX Runtime Web (JavaScript and WASM)1.22.0-dev.20250409-89f8206ba4MIT
    onnxruntime-common (inside Transformers.js)1.21.0MIT
    @huggingface/jinja (inside Transformers.js)0.5.3MIT

    Included voices: af_heart, af_bella, af_nicole, am_michael, am_fenrir, bf_emma. Browser bundles are fetched with pinned SHA-256 hashes; the eSpeak NG module is built from pinned source by scripts/build-espeak-ng.sh. offline-speech changes packaging/imports, adds an explicit phoneme-progress callback, and manages hosted voice caching.

    The full ONNX Runtime notices below include its upstream dependency inventory across runtime variants. They are a superset, not a claim that every listed native library ships in this browser build.

    Development and test dependencies (24 npm packages)

    These tools are not shipped to app visitors. Node.js 20 or later supplies build utilities; puppeteer-core, which needs Node.js 22.12 or later, and its complete locked dependency graph support browser tests. Tests use a separately installed Chrome/Chromium browser under that distribution’s own licenses.

    ComponentVersion / revisionLicense
    @puppeteer/browsers3.2.6Apache-2.0
    ansi-regex6.6.0MIT
    ansi-styles6.2.3MIT
    chromium-bidi157.0.8090-0Apache-2.0
    cliui9.0.1ISC
    devtools-protocol0.0.1704951BSD-3-Clause
    emoji-regex10.6.0MIT
    escalade3.2.0MIT
    get-caller-file2.0.5ISC
    get-east-asian-width1.7.0MIT
    mitt3.0.1MIT
    modern-tar0.8.6MIT
    puppeteer-core25.16.0Apache-2.0
    string-width7.2.0MIT
    string-width8.3.0MIT
    strip-ansi7.2.0MIT
    typed-query-selector2.12.3MIT
    webdriver-bidi-protocol1.0.1Apache-2.0
    wrap-ansi9.0.2MIT
    ws8.22.1MIT
    y18n5.0.8ISC
    yargs18.2.0MIT
    yargs-parser22.0.0ISC
    zod4.6.7MIT

    Full offline-speech license · Full third-party notices and license texts. Both files are saved with the page for offline reading.

    Corresponding source for the program and its libraries. offline-speech-source.tar.gz, espeak-ng-890071ff30231883d761e4d59edabd6d69faf159.tar.gz, kokoro-js-1.2.1-source.tar.gz, transformers-js-3.8.1-source.tar.gz, huggingface-jinja-0.5.3-source.tar.gz. The source index and source README give pinned source locations and build directions for every library, including ONNX Runtime and compiler/runtime components. This build is offline-speech commit a07797052d3e77916f403de2a0760def8d439976; its source is also published at https://github.com/dimacurrentai/offline-speech/tree/a07797052d3e77916f403de2a0760def8d439976.

    When redistributing this app, preserve notices and pass on the source archives and pinned source directions, or otherwise meet GPL corresponding-source requirements. A separate HTML page does not by itself determine the license obligations of a containing site. GNU GPL guidance on separate and combined works.