package edu.cmu.sphinx.frontend;
import java.util.Random;
import java.util.logging.Logger;
import org.testng.Assert;
import org.testng.annotations.Test;
import edu.cmu.sphinx.frontend.BaseDataProcessor;
import edu.cmu.sphinx.frontend.Data;
import edu.cmu.sphinx.frontend.DataProcessingException;
import edu.cmu.sphinx.frontend.DataStartSignal;
import edu.cmu.sphinx.frontend.DoubleData;
import edu.cmu.sphinx.frontend.FloatData;
import edu.cmu.sphinx.util.props.ConfigurationManager;
import edu.cmu.sphinx.util.props.PropertyException;
import edu.cmu.sphinx.util.props.PropertySheet;
import edu.cmu.sphinx.util.props.S4Boolean;
import edu.cmu.sphinx.util.props.S4Double;
import edu.cmu.sphinx.util.props.S4Integer;
/**
* A DataProcessor which inserts short speech snippets with variable length into a speech stream. The snippets are takes
* from the local neighborhood of the insertion point. The insertion behavior satisfies a uniform distribution with a
* user defined width. Additionally a insertion of can be enforced at the beginning of the stream.
*
* The purpose of this component is to randomize data to blow a small set of data up to a larger one.
* <p>
* This component should be put into the processing chain AFTER the SpeechClassifier or BEFORE a DataBlocker. The reason
* for that is, that the current implementation generates blocks with different lengths which would affect the results
* of the SpeechClassifier.
*
* @author Holger Brandl
*/
public class RandomSampleRepeater extends BaseDataProcessor {
/** The maximal value which could be added to the signal */
@S4Boolean(defaultValue = false)
public static final String PROP_RAND_STREAM_START = "forceRandStreamStart";
private boolean randStreamStart;
/**
* The maximal number of milliseconds which can be repeated using this randomizer. This number has to smaller than
* the length of the incoming data-blocks, because currently this class does not support data-buffering.
*/
@S4Double(defaultValue = 0.0)
public static final String PROP_MAX_REAPTED_MS = "maxRepeatedMs";
private double maxRepeatedMs;
private long maxRepeatedSamples = -1;
/**
* The property about the width of the uniform distribution which determines the distance between
* different repeated-sample-insertion-points.
*/
@S4Double(defaultValue = 100.0)
public static final String PROP_UNIFORM_DIST_WIDTH = "uDistWidthSec";
private double uDistWidthSec;
private long uDistWidthSamples = -1;
/**
* The property about using random seed or not for the randomization process. if not the stream
* will be randomized every time in the same manner.
*/
@S4Boolean(defaultValue = true)
public static final String PROP_USE_RANDSEED = "useRandSeed";
private boolean useRandSeed;
/** The property for the sample rate. */
@S4Integer(defaultValue = 16000)
public static final String PROP_SAMPLE_RATE = "sampleRate";
private int sampleRate;
/** The random generator used to compute the insertion points. */
private Random r;
private Logger logger;
private long nextInsertionPoint = 0;
private long numInsertedSamples = 0;
/*
* (non-Javadoc)
*
* @see edu.cmu.sphinx.frontend.DataProcessor#initialize(edu.cmu.sphinx.frontend.CommonConfig)
*/
@Override
public void initialize() {
super.initialize();
if (useRandSeed)
r = new Random();
else
r = new Random(12345);
uDistWidthSamples = RandomDataProcessor.ms2samples((int) (uDistWidthSec * 1000), sampleRate);
maxRepeatedSamples = RandomDataProcessor.ms2samples((int) (maxRepeatedMs), sampleRate);
}
private long computeNextInsertionPoint() {
// compute the next number in samples
int stepWidth = r.nextInt((int) uDistWidthSamples);
return stepWidth + nextInsertionPoint;
}
public void prepareForNewStream() {
if (randStreamStart)
nextInsertionPoint = 0;
else
nextInsertionPoint = computeNextInsertionPoint();
// reset the number of inserted samples
numInsertedSamples = 0;
}
/**
* Returns the next DoubleData object, which is a dithered version of the input
*
* @return the next available DoubleData object, or null if no Data is available
* @throws edu.cmu.sphinx.frontend.DataProcessingException
* if a data processing error occurred
*/
@Override
public Data getData() throws DataProcessingException {
Data input = getPredecessor().getData(); // get the spectrum
if (input instanceof DataStartSignal)
prepareForNewStream();
if (input != null && maxRepeatedMs != 0) {
if (input instanceof DoubleData || input instanceof FloatData) {
input = process(input);
}
}
return input;
}
/**
* Process data, adding dither
*
* @param input a MelSpectrum frame
* @return a mel Cepstrum frame
*/
private DoubleData process(Data input) throws IllegalArgumentException {
assert input instanceof DoubleData;
double[] inFeatures;
DoubleData doubleData = (DoubleData) input;
assert sampleRate == doubleData.getSampleRate();
inFeatures = doubleData.getValues();
// check wether we reached the next insertion point and repeat some parts of the input part if necessary
double[] extFeatures;
long firstSampleNumber = doubleData.getFirstSampleNumber() + numInsertedSamples;
long lastSampleNumber = doubleData.getFirstSampleNumber() + doubleData.getValues().length - 1;
if (nextInsertionPoint >= doubleData.getFirstSampleNumber() && nextInsertionPoint <= lastSampleNumber) {
int insertLength = Math.min(r.nextInt((int) maxRepeatedSamples) + 1, inFeatures.length);
// make sure that after insertion the block-length does not exceed 160 samples because with more SpeechClassifierNT will fail
assert doubleData.getValues().length + insertLength <= 160 : "block too large for next SpeechClassifier";
extFeatures = new double[insertLength + inFeatures.length];
logger.fine("RSR: repeat snippet with length " + insertLength + " at position " + nextInsertionPoint);
// copy the existing block into the new array and replicate the desired snippet inbetween
int startIndex = (int) (nextInsertionPoint - doubleData.getFirstSampleNumber());
System.arraycopy(inFeatures, 0, extFeatures, 0, startIndex);
System.arraycopy(inFeatures, 0, extFeatures, startIndex, insertLength);
System.arraycopy(inFeatures, startIndex, extFeatures, startIndex + insertLength, inFeatures.length - startIndex);
numInsertedSamples += insertLength;
nextInsertionPoint = computeNextInsertionPoint();
} else {
extFeatures = inFeatures;
}
DoubleData extendedData = new DoubleData(extFeatures, doubleData.getSampleRate(),
firstSampleNumber);
return extendedData;
}
/*
* (non-Javadoc)
*
* @see edu.cmu.sphinx.util.props.Configurable#newProperties(edu.cmu.sphinx.util.props.PropertySheet)
*/
@Override
public void newProperties(PropertySheet ps) throws PropertyException {
super.newProperties(ps);
logger = ps.getLogger();
maxRepeatedMs = ps.getDouble(RandomSampleRepeater.PROP_MAX_REAPTED_MS);
uDistWidthSec = ps.getDouble(RandomSampleRepeater.PROP_UNIFORM_DIST_WIDTH);
useRandSeed = ps.getBoolean(RandomSampleRepeater.PROP_USE_RANDSEED);
randStreamStart = ps.getBoolean(RandomSampleRepeater.PROP_RAND_STREAM_START);
sampleRate = ps.getInt(RandomSampleRepeater.PROP_SAMPLE_RATE);
initialize();
}
@Test
public void testInsertAtZero() {
RandomSampleRepeater rsr = ConfigurationManager.getInstance(RandomSampleRepeater.class);
rsr.randStreamStart = true;
rsr.useRandSeed = false;
rsr.initialize();
rsr.maxRepeatedSamples = 5;
rsr.numInsertedSamples = 20;
rsr.nextInsertionPoint = 5;
// create a dummy DoubleData
double[] data = new double[]{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
DoubleData dd = new DoubleData(data, 16000, 0);
DoubleData extData = rsr.process(dd);
double[] expectedData = new double[]{1, 2, 3, 4, 5, 1, 2, 6, 7, 8, 9, 10};
Assert.assertEquals(extData.getFirstSampleNumber(), 20, 0);
Assert.assertEquals(extData.getValues().length, expectedData.length);
for (int i = 0; i < expectedData.length; i++) {
Assert.assertEquals(extData.getValues()[i], expectedData[i], 0);
}
}
@Test
public void testInsertAt4() {
RandomSampleRepeater rsr = ConfigurationManager.getInstance(RandomSampleRepeater.class);
rsr.randStreamStart = true;
rsr.useRandSeed = false;
rsr.initialize();
rsr.maxRepeatedSamples = 3;
rsr.numInsertedSamples = 20;
rsr.nextInsertionPoint = 5;
// create a dummy DoubleData
double[] data = new double[]{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
DoubleData dd = new DoubleData(data, 16000, 0);
DoubleData extData = rsr.process(dd);
double[] expectedData = new double[]{1, 2, 3, 4, 5, 1, 2, 6, 7, 8, 9, 10};
Assert.assertEquals(extData.getFirstSampleNumber(), 20, 0);
Assert.assertEquals(extData.getValues().length, expectedData.length);
for (int i = 0; i < expectedData.length; i++) {
Assert.assertEquals(extData.getValues()[i], expectedData[i], 0);
}
}
@Test
public void testInsertAtBlockEnd() {
RandomSampleRepeater rsr = ConfigurationManager.getInstance(RandomSampleRepeater.class);
rsr.randStreamStart = true;
rsr.useRandSeed = false;
rsr.initialize();
rsr.maxRepeatedSamples = 3;
rsr.numInsertedSamples = 20;
rsr.nextInsertionPoint = 9;
// create a dummy DoubleData
double[] data = new double[]{1, 2, 3, 4, 5, 6, 7, 8, 9, 10};
DoubleData dd = new DoubleData(data, 16000, 0);
DoubleData extData = rsr.process(dd);
double[] expectedData = new double[]{1, 2, 3, 4, 5, 6, 7, 8, 9, 1, 2, 10};
Assert.assertEquals(extData.getFirstSampleNumber(), 20, 0);
Assert.assertEquals(extData.getValues().length, expectedData.length);
for (int i = 0; i < expectedData.length; i++) {
Assert.assertEquals(extData.getValues()[i], expectedData[i], 0);
}
}
}