Browser-recognizer

• A from-microphone speech recognizer built on Vosk that can be run on the browser, inspired by vosk-browser, but built from scratch and no code taken!
• Browser-recognizer can run both in the browser main thread and web workers.

Interface

• setLogLevel: set Kaldi's log level (default: -1)
  • -2: Error
  • -1: Warning
  • 0: Info
  • 1: Verbose
  • 2: More verbose
  • 3: Debug

Model and SpkModel

model = new Model()
spkModel = new SpkModel()
// Add events listeners
model.init(url, storepath, id)
spkModel.init(url, storepath, id)

Functions

• constructor: Construct the EventTarget part to enable addEventListener
• init : Initialize the internal object with an URL, storage path, and an ID.
  • If storepath contains valid model files and id is the same, there will not be a fetch from url.
  • If storepath doesn't contain valid model files, or if it contains valid model files but id is different, there will be a fetch from url, and the model is stored with id.
• delete: Delete self and free resources

Events

• ready: The model is ready to be put into a recognizer via the constructor, or setSpkModel() for SpkModel.
• error: An error occured, check the event's "details" property.

Recognizer

recognizer = new Recognizer()
// Add event listeners
recognizer.init(model)

Functions

• constructor: Construct the EventTarget part to enable addEventListener
• init: Construct the real internal object from a model
• start: Start recognizing
• stop: Stop recognizing
• setWords: Return words' information in a result event (default: false)
• setPartialWords: Return words' information in a partialResult event (default: false)
• setNLSML: Return result and partialResult in NLSML form (default: false)
• setMaxAlternatives: Set the max number of alternatives for result event (default: false)
• setGrm: Add grammar to the recognizer (default: none)
• setSpkModel: Set the speaker model of the recognizer (default: none)
• delete: Call stop, delete self and free all resource

Events

• partialResult: There is a partial recognition result, check the event's "details" property
• result: There is a full recognition result, check the event's "details" property
• error: An error occured, check the event's "details" property.

Other key points

IMPORTANT

• You MUST call delete() on objects at the end of its usage. Or put:

  __genericObj__.objects.forEach(obj => obj.delete())
  

  at the end of your program to automatically do that. We have to do this because Emscripten doesn't call destructors. See here.

• To be safe, always handle the API through events by adding all event listener before calling init() on objects.

• Always call init on the regular Model object before calling init on the recognizer. SpkModel can be init and set later.

Guarantees

• If an error occurs (error event is fired), no changes was made, and no other dependent events will fire. For example, if an error occur while loading the model, the "ready" event won't fire in order to prevent executing code on a nonexistent model.

Limitations compared to vosk-browser:

• Microphone only
• Fixed memory size at 300MB, changing it require recompilation

Additions to vosk-browser:

• Multiple models support
• Speaker model (SpkModel) support
• Storage path management (when many models are required)
• Model ID management (when model updates are required)

This requires SharedArrayBuffer, so set the response headers:

• Cross-Origin-Embedder-Policy ---> require-corp
• Cross-Origin-Opener-Policy ---> same-origin

If you can't set these headers, you can use a VERY HACKY workaround at src/addCOI.js.

Usage

<!--Load this from a script tag-->
<script src="BrowserRecognizer.js"></script>
<!-->
<script>

</script>