pdg-react-native-vosk

react-native-vosk - React ASR (Automated Speech Recognition)

Speech recognition module for react native using Vosk library

Installation

Library

npm install -S react-native-vosk

Models

Vosk uses prebuilt models to perform speech recognition offline. You have to download the model(s) that you need on Vosk official website Avoid using too heavy models, because the computation time required to load them into your app could lead to bad user experience. Then, unzip the model in your app folder. If you just need to use the iOS version, put the model folder wherever you want, and import it as described below. If you need both iOS and Android to work, you can avoid to copy the model twice for both projects by importing the model from the Android assets folder in XCode.

Experimental: Loading a model dynamically into the app storage, aside from the main bundle is a new and experimental feature. Would love for you all to test, and let us know if it is a viable option. If you choose to download a model to your app’s storage (preferably internal), you can pass the model directory path when calling vosk.loadModel(path).

To download and load a model as part of an app's Main Bundle, just do as follows:

Android

In Android Studio, open the project manager, right-click on your project folder and go to New > Folder > Assets folder.

Then put the model folder inside the assets folder created. In your file tree it should be located in android\app\src\main\assets. So, if you downloaded the french model named model-fr-fr, you should access the model by going to android\app\src\main\assets\model-fr-fr. In Android studio, your project structure should be like that:

You can import as many models as you want.

iOS

In XCode, right-click on your project folder, and click on "Add files to [your project name]".

Then navigate to your model folder. You can navigate to your Android assets folder as mentionned before, and chose your model here. It will avoid to have the model copied twice in your project. If you don't use the Android build, you can just put the model wherever you want, and select it.

That's all. The model folder should appear in your project. When you click on it, your project target should be checked (see below).

Usage

import Vosk from 'react-native-vosk';

// ...

const vosk = new Vosk();

vosk
  .loadModel('model-en-en')
  .then(() => {
    const options = {
      grammar: ['left', 'right', '[unk]'],
    };

    vosk
      .start(options)
      .then(() => {
        console.log('Recognizer successfuly started');
      })
      .catch((e) => {
        console.log('Error: ' + e);
      });

    const resultEvent = vosk.onResult((res) => {
      console.log('A onResult event has been caught: ' + res);
    });

    // Don't forget to call resultEvent.remove(); to delete the listener
  })
  .catch((e) => {
    console.error(e);
  });

Note that start() method will ask for audio record permission.

See complete example...

Experimental Loading via Path

• Primarily intended for models that are not included in the app’s Main Bundle.

Preliminary Steps

• Use a file system package to download and store a model from remote location
  • react-native-file-access is one that we found to be stable, but this is a personal preference based on use

import Vosk from 'react-native-vosk';

// ...

const vosk = new Vosk();

const path = "some/path/to/model/directory";

vosk
  .loadModel(path)
  .then(() => {
    const options = {
      grammar: ['left', 'right', '[unk]'],
    };

    vosk
      .start(options)
      .then(() => {
        console.log('Recognizer successfuly started');
      })
      .catch((e) => {
        console.log('Error: ' + e);
      });

    const resultEvent = vosk.onResult((res) => {
      console.log('A onResult event has been caught: ' + res);
    });

    // Don't forget to call resultEvent.remove(); to delete the listener
  })
  .catch((e) => {
    console.error(e);
  });

Methods

Method	Argument	Return	Description
loadModel	path: string	Promise<void>	Loads the voice model used for recognition, it is required before using start method.
start	options: VoskOptions or none	Promise<void>	Starts the recognizer, an onResult() event will be fired.
stop	none	none	Stops the recognizer. Listener should receive final result if there is any.
unload	none	none	Unloads the model, also stops the recognizer.

Types

VoskOptions	Type	Required	Description
grammar	string[]	No	Set of phrases the recognizer will seek on which is the closest one from the record, add "[unk]" to the set to recognize phrases striclty.
timeout	int	No	Timeout in milliseconds to listen.

Events

Method	Promise return	Description
onPartialResult	The recognized word as a string	Called when partial recognition result is available.
onResult	The recognized word as a string	Called after silence occured.
onFinalResult	The recognized word as a string	Called after stream end, like a stop() call
onError	The error that occured as a string or exception	Called when an error occurs
onTimeout	void	Called after timeout expired

Examples

Default

vosk.start().then(() => {
  const resultEvent = vosk.onResult((res) => {
    console.log('A onResult event has been caught: ' + res);
  });
});

// when done, remember to call resultEvent.remove();

Using grammar

vosk.start({
  grammar: ['left', 'right', '[unk]'],
}).then(() => {
  const resultEvent = vosk.onResult((res) => {
    if (res === 'left') {
      console.log('Go left');
    } else if (res === 'right') {
      console.log('Go right');
    } else {
      console.log("Instruction couldn't be recognized");
    }
  });
});

// when done, remember to call resultEvent.remove();

Using timeout

vosk.start({
  timeout: 5000,
}).then(() => {
  const resultEvent = vosk.onResult((res) => {
    console.log('An onResult event has been caught: ' + res);
  });

  const timeoutEvent = vosk.onTimeout(() => {
    console.log('Recognizer timed out');
  });
})

// when done, remember to clean all listeners;

Complete example

Contributing

See the contributing guide to learn how to contribute to the repository and the development workflow.

License

MIT