How to Extract MFCC features in Java

Question

I am working on converting a speech recognition project from MATLAB to Java code. I have been able to read the .wav files (as vectors of values in the range -1 to 1) using the java example provided here. This works exactly as the wavread function in MATLAB.

My next task is to extract the MFCC feature vectors from the original sample vectors. In MATLAB, I used Voicebox to achieve this easily but have been unable to find a Java equivalent. Using Voicebox, I have a code like:

a = melcepst(samples(1,:), 44100)

where each row of 'samples' contains the vector representing each .wav sample. The method returns a 2-dimensional matrix of the MFCC features for each sample.

I have seen Sphinx, but have been unable to understand how to use it for this task. Any help with using Sphinx or any other java solution will be appreciated.

To extract MFCC, FTT values from any audio file in Java, recently I have implemented java program. This program generates Librosa equivalent values of MFCC and FTT. As its not small piece of code so cant post complete code snippet here for quick glance. https://github.com/Subtitle-Synchronizer/jlibrosa — Abhishek, Aug 11 '20 at 17:47

score 5 · Answer 1 · answered Aug 13 '15 at 22:11

You can get MFCC frames using Sphinx 4 doing something like this:

AudioFileDataSource audioDataSource = new AudioFileDataSource(3200, null);
audioDataSource.setAudioFile(new URL("file:///path/to/my.wav", "source");

final ArrayList<DataProcessor> pipeline = new ArrayList<DataProcessor>();

pipeline.add(audioSource);
pipeline.add(new DiscreteFourierTransform());
pipeline.add(new MelFrequencyFilterBank(minFreq, maxFreq, numFilters));
pipeline.add(new DiscreteCosineTransform2(numFilters, 12));
FrontEnd f = new FrontEnd(pipeline);

Data mfccs;
do {
    mfccs = f.getData();
} while(mfccs != null);

How to Extract MFCC features in Java

1 Answers1