Transcribe.js

Transcribe speech to text in the browser. Based on a wasm build of whisper.cpp.

Note: This package is browser only. Node.js is not supported. (see this discussion for details)

Packages

All packages are under @transcribe namespace.

Package	Description
@transcribe/shout	Wasm build based on whisper.cpp. Contains Module file including the wasm binary and a separate webworker file.
@transcribe/transcriber	`FileTranscriber` and `StreamTranscriber` for transcribing media files or streams.

Prerequisite

Webserver

Your webserver must serve the files with cross origin headers.

"Cross-Origin-Embedder-Policy": "require-corp"
"Cross-Origin-Opener-Policy": "same-origin"

Browser

Your browser must support SharedArrayBuffer. (brower support)

The default wasm files are built with SIMD enabled. If your browser/device doens't support SIMD use the no-simd files instead. Also check out the example code on how to use it. (brower support)

Model File

You need a ggml model file to run Transcribe.js. You can download them on hugging face https://huggingface.co/ggerganov/whisper.cpp/tree/main . You should start with the (quantized) tiny or base models. Larger models propably won't work but you can try it, though.

Installation

Svelte
SvelteKit
Vue
Angular
Other

NPM (without bundler)

Install shout wasm and transcriber packages

npm install --save @transcribe/transcriber

The shout.wasm files must be accessable and served by your webserver. Depending on your project setup you may need to copy them from node_modules to your public directory.

# copy shout wasm
cp node_modules/@transcribe/shout/src/shout/shout.wasm.worker.mjs /your/project
cp node_modules/@transcribe/shout/src/shout/shout.wasm.js /your/project

# optional: copy no-simd build
cp node_modules/@transcribe/shout/src/shout/shout.wasm.worker_no-simd.mjs /your/project
cp node_modules/@transcribe/shout/src/shout/shout.wasm_no-simd.js /your/project

# optional: copy audio-worklets, only needed if you want to use StreamTranscriber
cp -r node_modules/@transcribe/transcriber/src/audio-worklets /your/project

Manual Installation

You can use Transcribe.js without a bundler or package manager. Download the files from this repository, copy the src/* directories to your webserver and include the following into your HTML. Make sure to set the correct paths in the import map.

<!-- set paths to js files -->
<script type="importmap">
  {
    "imports": {
      "@transcribe/shout": "/src/shout/shout.wasm.js",
      "@transcribe/transcriber": "/src/index.js"
    }
  }
</script>

<!-- use type="module" for es6 imports -->
<script type="module">
  import createModule from "/your/project/shout.wasm.js"; // path where you've copied before
  // import createModule from "@transcribe/shout";  // if you use an import map
  import { FileTranscriber } from "@transcribe/transcriber";

  ...
</script>

Usage

For full code examples and advanced usage please see https://www.transcribejs.dev or check out the File Transcriber Example.

import createModule from "@transcribe/shout"; // if you use import map or bundler like vite
// import createModule from "/your/project/shout.wasm.js"; // you can also exclude @transcibe/shout from your bundler and import manually
import { FileTranscriber } from "@transcribe/transcriber";

// create new instance
const transcriber = new FileTranscriber({
  createModule, // create module function from emscripten build
  model: "/your/project/ggml-tiny-q5_1.bin", // path to ggml model file
  // workerPath: "/your/project", // only set if you don't use a bundler; directory of shout.wasm.worker.mjs copied before
});

// init wasm transcriber worker
await transcriber.init();

// transcribe audio/video file
const result = await transcriber.transcribe("/your/project/my.mp3");

console.log(result);

The result is an JSON object containg the text segements and timestamps.

{
  "result": {
    "language": "en"
  },
  "transcription": [
    {
      "timestamps": {
        "from": "00:00:00,000",
        "to": "00:00:11,000"
      },
      "offsets": {
        "from": 0,
        "to": 11000
      },
      "text": " And so my fellow Americans ask not what your country can do for you, ask what you can do for your country.",
      "tokens": [
        {
          "text": " And",
          "timestamps": {
            "from": "00:00:00,320",
            "to": "00:00:00,350"
          },
          "offsets": {
            "from": 320,
            "to": 350
          },
          "id": 400,
          "p": 0.726615 // propability, aka. how likely the estimate is true, 0..1, 1 is best
        },
        // ... one token per word
      ]
    }
  ]
}

Development

Install Emscripten and its required tools.

Clone the repository, install dependencies, start the dev server and open http://localhost:9876/examples/index.html in your browser.

git clone https://github/transcribejs/transcribe.js
cd transcribe
npm install
npm run dev

Types

The library is not written in typescript. This way no extra build step is needed during development and in production.

To still get proper type support type definitions get generated from JSDoc comments.

npm run generate-types

Wasm build

The whisper.cpp repository is a git submodule. To get the latest version of whisper.cpp go into the directory and pull the latest changes from github.

cd shout.wasm/whisper.cpp
git pull origin master

The wasm files are build from shout.wasm/src/shout.wasm.cpp. If you want to add new functions from whisper.cpp to the wasm build this is the file to add them.

I'm pretty sure that this will not compile on every machine/architecture, but I'm no expert in C++. If you know how to optimize the build process please let me know or create a pull request. Maybe this should be dockerized.?

# run cmake to build wasm
npm run wasm:build

# copy emscripten build files to project
npm run wasm:copy

Tests

Unit/functional tests for the Transcriber functions.

npm run test:unit

E2E tests using Playwright. Firefox somehow needs waaaaaay longer during e2e test than in a the "real" browser.

npm run test:e2e

or use the Playwright UI for details

npm run test:e2e-ui

Credits

People

Many thanks to the people who supported this project, be it through code, ideas or general testing. I appreciate your time and effort.

@MarketingPip - testing on older devices

Libraries

Also thank you to the creators and contributors of the following open source libraries that were used in this project:

whisper.cpp: A C++ implementation of whisper. GitHub Repository
emscripten: A toolchain for compiling C and C++ code to WebAssembly. Official Site
water.css: A minimal CSS framework for styling HTML. Official Site
fft.js: A library for Fast Fourier Transform calculations. GitHub Repository
Moattar, Mohammad & Homayoonpoor, Mahdi. (2010). A simple but efficient real-time voice activity detection algorithm. Research Paper
vitest: A website for testing voice recognition. Official Site
Playwright: A tool for automating browser testing. Official Site

Audio Test Files

examples/albert.ogg Radio Universidad Nacional de La Plata, CC BY-SA 3.0, via Wikimedia Commons
examples/jfk.wav: CC BY-SA 3.0, via Wikimedia Commons

Sponsoring

This project is tested with BrowserStack

@transcribe/transcriber

Transcribe.js

Packages

Prerequisite

Webserver

Browser

Model File

Installation

NPM (without bundler)

Manual Installation

Usage

Development

Types

Wasm build

Tests

Credits

People

Libraries

Audio Test Files

Sponsoring

Readme

Keywords

Package Sidebar

Install

Repository

Homepage

Weekly Downloads

Version

License

Unpacked Size

Total Files

Last publish

Collaborators

@transcribe/transcriber

Transcribe.js

Packages

Prerequisite

Webserver

Browser

Model File

Installation

NPM (without bundler)

Manual Installation

Usage

Development

Types

Wasm build

Tests

Credits

People

Libraries

Audio Test Files

Sponsoring

Readme

Keywords

Package Sidebar

Install

Repository

Homepage

DownloadsWeekly Downloads

Version

License

Unpacked Size

Total Files

Last publish

Collaborators

Weekly Downloads