CMU Sphinx Transcription accuracy

Question

I am working on a web speech recognition app. I am using recorderJS to capture the sound and send it to the backend where it should be processed using CMU Sphinx. I've had accuracy problems while discovering the library with the latest verion 5prealpha using the default acoustic Model , langage model and dictionnary and later reducing the number of recognized words by using a JSGF grammar so I used the 1.0 beta6 version. The microphone recognition with the 1.0 beta6 version is pretty accurate however when I transcribe the sound it's always poor.I am wondering how to improve the accuracy.I tried using the StreamSpeechRecognizer with the latest version but it also gives poor results.

To get help on the accuracy you need to provide an audio recording you want to convert. If you work with JS, it's better to use something like dictate.js — Nikolay Shmyrev

user3024469 user3024469 · Accepted Answer · 2016-08-03T09:03:51

I managed to get a good accuracy. I checked the implementation of the edu.cmu.sphinx.frontend.util.Microphone class and I found out that the sample rate was 16000 , the bit rate is 16 and the number of channels equals 1. I looked further in the recorderJS and I found out that in Google chrome the sample rate was 44100 so I looked for a configurable version of the library and I found Chris Rudmin fork of Matt Diamond's RecorderJS. I didn't use the latest version because the sound is exported in ogg and I need it to be wav so I looked in previous releases ; I used version 0.3 where the bit rate is configurable and it worked fine. I later modified the example that comes with it and the following parameters gave a good accuracy:

monitor gain : 0

bitDepth : 16

number of channels : 1

recordOpus : unchecked

sample rate : 16000

bit rate : 32000

This is the config of the stream data source in the cmu sphinx's xml config file.

    type="edu.cmu.sphinx.frontend.util.StreamDataSource">

    <property name="sampleRate" value="16000" />

    <property name="bitsPerSample" value="16" />

    <property name="bigEndianData" value="false" />

    <property name="signedData" value="true" />

</component>

CMU Sphinx Transcription accuracy

1 Answers