PocketSphinx for Node.js

A Node.js Binding for PocketSphinx

This module aims to allow basic speech recognition on portable devices through the use of PocketSphinx. It is possible to either use buffer chunks for live recognition or pass audio buffers with the complete audio to PocketSphinx to decode.

Prerequisites

To build the node extension you have to install Sphinxbase and PocketSphinx from the repositories preferably:

• Sphinxbase repository
• PocketSphinx repository

Example

var PocketSphinx = require('pocketsphinx');

var ps = new PocketSphinx.Recognizer({
	'-hmm': '/file/path',
	'-dict': '/file/path',
	'-nfft': 512,
	'-remove_silence': false	
});

ps.on('hyp', function(err, hypothesis, score){
	if(err) console.error(err);
	console.log('Hypothesis: ', hypothesis);
});

ps.addKeyphraseSearch("keyphrase_name", "keyphrase");
ps.addKeywordSearch("keyword_name", "/file/path");
ps.addGrammarSearch("grammar_name", "/file/path");
ps.addNgramSearch("ngram_name", "/file/path");

ps.search = "keyphrase_name";

ps.start();
ps.write(data);
ps.stop();

Options

You can pass any valid PocketSphinx parameter you could e.g. pass to pocketsphinx_continuous but you have to assure you're using the correct type for the parameter. If you are for example pass '-samprate' as a string like '44100.0' the configuration will fail. In the following list you can see some of the probably most required options with their required data type:

option	type	default	description
-samprate	float	44100.0	The sample rate of the passed data
-hmm	string	modelDirectory + "/en-us/en-us"	The hmm model directory
-dict	string	modelDirectory + "/en-us/cmudict-en-us.dict"	The dictionary file directory
-nfft	int	2048	The nfft value
For more options you can look into the manual of pocketsphinx_continuous with $ man pocketsphinx_continuous

Methods

The PocketSphinx Object itself has the properties

• Recognizer(options, [hyp]) - Creates a new Recognizer instance
• modelDirectory - The default model directory
• fromFloat(buffer):buffer - Resamples javascript audio buffers to use with PocketSphinx

A Recognizer instance has the following methods:

• on(event, function) - Attaches an event handler (overwrites old event handlers for this event)
• off(event) - Removes an event handler
• start() - Starts the decoder
• stop() - Stops the decoder
• restart() - Restarts the decoder
• reconfig(options, [hyp]) - Reconfigures the decoder without having to reload it
• silenceDetection(enabled) - Disables or enables silence detection (Default: enabled)
• addKeyphraseSearch(name, keyphrase) - Adds a keyphrase search
• addKeywordsSearch(name, keywordFile) - Adds a keyword search
• addGrammarSearch(name, jsgfFile) - Adds a jsgf search
• addNgramSearch(name, nGramFile) - Adds a nGram search
• write(buffer) - Decodes a complete audio buffer
• writeSync(buffer) - Decodes the next audio buffer chunk
• lookupWords(array):object - Returns an object with the properties in (an object with words in dictionary and their phonetic transcription as value) and out (an array with out of dictionary words)
• addWords(object) - Adds the phonetic transcription from object to dictionary (key = word, value = transcription)
• free() - Releases all resources associated with the decoder.

Events

The following events are currently supported

event	parameters	description
hyp	error, hypothesis, score	When a hypothesis is available. score is the path score corresponding to returned string.
hypFinal	error, hypothesis, isFinal	When a hypothesis is available. isFinal indicates if hypothesis has reached final state in the grammar.
start	none	When decoding started.
stop	none	When decoding stopped.
speechDetected	none	When speech was detected the first time.
silenceDetected	none	When silence was detected after speech.
error	error	When an error occurred

Specify a search

To specify a search you can use one of the add functions mentioned in the methods section above and then add the name to the instance's search accessor like so:

ps.addGrammarSearch('someSearch', 'digits.gram');
ps.search = 'someSearch';
ps.start();

Or you can pass e.g. a language model file, a jsgf grammar or a keyword file directly with the Recognizer options or reconfigure the Recognizer with such options at runtime:

var ps = new PocketSphinx.Recognizer({
	'-lm': '/path/myFancyLanguageModel' // This adds an nGram search
});
ps.on('hyp', function(err, hypothesis, score) {
	if(err) console.error(err);
	console.log('Hypothesis: ', hypothesis);
});
ps.start();