Leopard Speech-to-Text engine for web browsers (via WebAssembly)
npm install @picovoice/leopard-webMade in Vancouver, Canada by Picovoice
Leopard is an on-device speech-to-text engine. Leopard is:
- Private; All voice processing runs locally.
- Accurate
- Compact and Computationally-Efficient
- Cross-Platform:
- Linux (x86_64), macOS (x86_64, arm64), Windows (x86_64)
- Android and iOS
- Chrome, Safari, Firefox, and Edge
- Raspberry Pi (3, 4, 5)
- Chrome / Edge
- Firefox
- Safari
The Leopard Web Binding uses SharedArrayBuffer.
Include the following headers in the response to enable the use of SharedArrayBuffers:
```
Cross-Origin-Opener-Policy: same-origin
Cross-Origin-Embedder-Policy: require-corp
Refer to our Web demo for an example on creating a server with the corresponding response headers.
Browsers that don't support SharedArrayBuffers or applications that don't include the required headers will fall back to using standard ArrayBuffers. This will disable multithreaded processing.
IndexedDB is required to use Leopard in a worker thread. Browsers without IndexedDB supportLeopard
(i.e. Firefox Incognito Mode) should use in the main thread.
Multi-threading is only enabled for Leopard when using on a web worker.
Using yarn:
`console`
yarn add @picovoice/leopard-web
or using npm:
`console`
npm install --save @picovoice/leopard-web
Leopard requires a valid Picovoice AccessKey at initialization. AccessKey acts as your credentials when using Leopard SDKs.AccessKey
You can get your for free. Make sure to keep your AccessKey secret.AccessKey
Signup or Login to Picovoice Console to get your .
Create a model in Picovoice Console or use one of the default language models found in lib/common.
For the web packages, there are two methods to initialize Leopard.
NOTE: Due to modern browser limitations of using a file URL, this method does __not__ work if used without hosting a server.
This method fetches the model file from the public directory and feeds it to Leopard. Copy the model file into the public directory:
`console`
cp ${LEOPARD_MODEL_FILE} ${PATH_TO_PUBLIC_DIRECTORY}
NOTE: This method works without hosting a server, but increases the size of the model file roughly by 33%.
This method uses a base64 string of the model file and feeds it to Leopard. Use the built-in script pvbase64 to
base64 your model file:
`console`
npx pvbase64 -i ${LEOPARD_MODEL_FILE} -o ${OUTPUT_DIRECTORY}/${MODEL_NAME}.js
The output will be a js file which you can import into any file of your project. For detailed information about pvbase64,
run:
`console`
npx pvbase64 -h
Leopard saves and caches your model file in IndexedDB to be used by WebAssembly. Use a different customWritePath variableforceWrite
to hold multiple models and set the value to true to force re-save a model file.
Either base64 or publicPath must be set to instantiate Leopard. If both are set, Leopard will use the base64 model.
`typescript
const leopardModel = {
publicPath: ${MODEL_RELATIVE_PATH},
// or
base64: ${MODEL_BASE64_STRING},
// Optionals
customWritePath: "leopard_model",
forceWrite: false,
version: 1,
}
`
Create an instance of Leopard in the main thread:
`typescript`
const leopard = await Leopard.create(
${ACCESS_KEY},
leopardModel,
options
);
Or create an instance of Leopard in a worker thread:
`typescript`
const leopard = await LeopardWorker.create(
${ACCESS_KEY},
leopardModel,
options
);
Additional configuration options can be passed to create. Set enableAutomaticPunctuation to true if you wish to enable punctuation in transcript or enableDiarization if you wish to enable speaker diarization.
`typescript`
const options = {
enableAutomaticPunctuation: true,
enableDiarization: true
}
The process result is an object with:
- transcript: A string containing the transcribed data.words
- : A list of objects containing a word, startSec, endSec, confidence and speakerTag.
`typescript
function getAudioData(): Int16Array {
... // function to get audio data
return new Int16Array();
}
const result = await leopard.process(getAudioData());
console.log(result.transcript);
console.log(result.words);
`
For processing using worker, you may consider transferring the buffer instead for performance:
`typescript`
let pcm = new Int16Array();
const result = await leopard.process(pcm, {
transfer: true,
transferCallback: (data) => { pcm = data }
});
console.log(result.transcript);
console.log(result.words);
Clean up used resources by Leopard or LeopardWorker:
`typescript`
await leopard.release();
Terminate LeopardWorker instance:
`typescript`
await leopard.terminate();
Along with the transcript, Leopard returns metadata for each transcribed word. Available metadata items are:
- Start Time: Indicates when the word started in the transcribed audio. Value is in seconds.
- End Time: Indicates when the word ended in the transcribed audio. Value is in seconds.
- Confidence: Leopard's confidence that the transcribed word is accurate. It is a number within [0, 1].0
- Speaker Tag: If speaker diarization is enabled on initialization, the speaker tag is a non-negative integer identifying unique speakers, with reserved for unknown speakers. If speaker diarization is not enabled, the value will always be -1`.
For example usage refer to our Web demo application.