mfcc.zip_MFCC_mfcc java_特征提取 java

/**Calculates the mel-based cepstra coefficients for one frame of speech. * Based on the original MFCC implementation described in: * [1] Davis & Mermelstein - IEEE Transactions on ASSP, August 1980. * Additional references are: * [2] Joseph Picone, Proceedings of the IEEE, Sep. 1993. * [3] Jankowski et al. IEEE Trans. on Speech and Audio Processing. July, 1995. * [4] Cardin et al, ICASSP'93 - pp. II-243 * * Notice that there are several different implementations of the mel filter * bank. For example, the log is usually implementated after having the filter * outputs calculated, but could be implemented before filtering. Besides, there are * differences in the specification of the filter frequencies. [1] * suggested linear scale until 1000 Hz and logarithm scale afterwards. * This implementation uses the equation (10) in [2]: * mel frequency = 2595 log(1 + (f/700)), where log is base 10 * to find the filter bank center frequencies. * * @author Aldebaro Klautau * @version 2.0 - March 07, 2001 * @see MFCCPatternGenerator */ //if m_oisZeroThCepstralCoefficientCalculated is true, //this class decrements m_nnumberOfParameters by 1 and //adds the 0-th coefficient to complete a vector with //the number of MFCC's specified by the user. public class MFCC { // parameter USEPOWER in HTK, where default is false private static final boolean m_ousePowerInsteadOfMagnitude = false; // Number of MFCCs per speech frame. //每帧信号的MFCC维数 private final int m_nnumberOfParameters; /** * Sampling frequency. */ //采样率 private final double m_dsamplingFrequency; /** * Number of filter in mel filter bank. */ //梅尔滤波器的个数 private final int m_nnumberOfFilters; /** * Number of FFT points. */ //每次FFT变换的点数 private final int m_nFFTLength; /** * Coefficient of filtering performing in cepstral domain (called * 'liftering' operation). It is not used if m_oisLifteringEnabled is false. */ //如果m_ois。。是false，就用不到 private final int m_nlifteringCoefficient; /** * True enables liftering. */ //一个允许“同态滤波参数”的标志位 private final boolean m_oisLifteringEnabled; /** * Minimum value of filter output, otherwise the log is not calculated and * m_dlogFilterOutputFloor is adopted. ISIP implementation assumes * m_dminimumFilterOutput = 1 and this value is used here. */ //滤波器输出的最小值 private final double m_dminimumFilterOutput = 1.0; /** * True if the zero'th MFCC should be calculated. */ //设置对数能量是否应该被计算 private final boolean m_oisZeroThCepstralCoefficientCalculated; /** * Floor value for filter output in log domain. ISIP implementation assumes * m_dlogFilterOutputFloor = 0 and this value is used here. */ //滤波器在对数域输出的最低限值 private final double m_dlogFilterOutputFloor = 0.0; private int[][] m_nboundariesDFTBins; private double[][] m_dweights; private FFT m_fft; private double[][] m_ddCTMatrix; private double[] m_dfilterOutput; private final double[] m_nlifteringMultiplicationFactor; // things to be calculated just once: private final double m_dscalingFactor; /** * The 0-th coefficient is included in nnumberOfParameters. So, if one wants * 12 MFCC's and additionally the 0-th coefficient, one should call the * constructor with nnumberOfParameters = 13 and * oisZeroThCepstralCoefficientCalculated = true */ //对数能量也可以包括在n维的参数中，所以如果你想要12维MFCC外加对数能量参数，你应该在构造函数中设参数为13维，并且设置短时能量计算为True // 构造函数中各种初始化分配内存等操作 public MFCC(int nnumberOfParameters, double dsamplingFrequency, int nnumberofFilters, int nFFTLength, boolean oisLifteringEnabled, int nlifteringCoefficient, boolean oisZeroThCepstralCoefficientCalculated) { m_oisZeroThCepstralCoefficientCalculated = oisZeroThCepstralCoefficientCalculated; if (m_oisZeroThCepstralCoefficientCalculated) { // the user shouldn't notice that nnumberOfParameters was // decremented internally m_nnumberOfParameters = nnumberOfParameters - 1; } else { m_nnumberOfParameters = nnumberOfParameters; } m_dsamplingFrequency = dsamplingFrequency; m_nnumberOfFilters = nnumberofFilters; m_nFFTLength = nFFTLength; // the filter bank weights, FFT's cosines and sines // and DCT matrix are initialized once to save computations. //初始化各种参数来减少计算量 // initializes the mel-based filter bank structure //初始化梅尔三角滤波器结构 calculateMelBasedFilterBank(dsamplingFrequency, nnumberofFilters, nFFTLength); m_fft = new FFT(m_nFFTLength); // initialize FFT 初始化FFT initializeDCTMatrix();//初始化DCT矩阵 m_nlifteringCoefficient = nlifteringCoefficient; m_oisLifteringEnabled = oisLifteringEnabled; // avoid allocating RAM space repeatedly, m_dfilterOutput is // going to be used in method getParameters() 避免分开分配内存，m_dfilterOutput数组将会在getParametrers()中用到 m_dfilterOutput = new double[m_nnumberOfFilters]; //滤波器输出数组 // needed in method getParameters() // m_dscalingFactor shouldn't be necessary because it's only // a scaling(缩放) factor, but I'll implement it // for the sake of(为了) getting the same numbers ISIP gets m_dscalingFactor = Math.sqrt(2.0 / m_nnumberOfFilters); // for liftering method if (m_oisLifteringEnabled) { // note that: @SuppressWarnings("unused") int nnumberOfCoefficientsToLift = m_nnumberOfParameters; // even when m_oisZeroThCepstralCoefficientCalculated is true // because if 0-th cepstral coefficient is included, // it is not liftered m_nlifteringMultiplicationFactor = new double[m_nlifteringCoefficient]; double dfactor = m_nlifteringCoefficient / 2.0; double dfactor2 = Math.PI / m_nlifteringCoefficient; for (int i = 0; i < m_nlifteringCoefficient; i++) { m_nlifteringMultiplicationFactor[i] = 1.0 + dfactor * Math.sin(dfactor2 * (i + 1)); } if (m_nnumberOfParameters > m_nlifteringCoefficient) { new Error( "Liftering is enabled and the number " + "of parameters = " + m_nnumberOfParameters + ", while " + "the liftering coefficient is " + m_nlifteringCoefficient + ". In this case some cepstrum coefficients would be made " + "equal to zero due to liftering, what does not make much " + "sense in a speech recognition system. You may want to " + "increase the liftering coefficient or decrease the number " + "of MFCC parameters."); } } else { m_nlifteringMultiplicationFactor = null; } } /** Initializes the DCT matrix. */ private void initializeDCTMatrix() { m_ddCTMatrix = new double[m_nnumberOfParameters][m_nnumberOfFilters]; for (int i = 0; i < m_nnumberOfParameters; i++) { for (int j = 0; j < m_nnumberOfFilters; j++) { m_ddCTMatrix[i][j] = Math.cos((i + 1.0) * (j + 1.0 - 0.5) //此处因为i和j都是从0到m_nnumber-1计算的，而正规公式是1到m_nnumber计算的，所以计算中有“i+1”和“j+1”这一块 * (Math.PI / m_nnumberOfFilters)); } } } /** * Converts frequencies in Hz to mel scale according to mel frequency = 2595 * log(1 + (f/700)), where log is base 10 and f is the frequency in Hz. */ //将频率转换为梅尔频率，此处采样率并没有用到，不知道为什么要加这个参数 public static double[] convertHzToMel(double[] dhzFrequencies, double dsamplingFrequency) { double[] dmelFrequencies = new double[dhzFrequencies.length]; for (int k = 0; k < dhzFrequencies.length; k++) { dmelFrequencies[k] = 2595.0 * (Math .log(1.0 + (dhzFrequencies[k] / 700.0)) / Math.log(10)); } return dmelFrequencies; } /** * Calculates triangular filters. 三角带通滤波器（Triangular Bandpass * Filters）：将能量频谱能量乘以一组 20 个三角带通滤波器�