mfcc

Functions

int main(int argc, char *argv[])

mfcc [ option ] [ infile ]

• -n int

  • number of channels \((1 \le C)\)

• -m int

  • order of coeffcients \((1 \le M)\)

• -l int

  • FFT length \((2 \le N)\)

• -c int

  • liftering parameter \((1 \le L)\)

• -s double

  • sampling rate in kHz \((0 < F_s)\)

• -L double

  • lowest frequency in Hz \((0 \le F_l < F_h)\)

• -H double

  • highest frequency in Hz \((F_l < F_h \le 500F_s)\)

• -q int

  • input format

    • 0 amplitude spectrum in dB

    • 1 log amplitude spectrum

    • 2 amplitude spectrum

    • 3 power spectrum

    • 4 windowed waveform

• -o int

  • output format

    • 0 MFCC

    • 1 MFCC and energy

    • 2 MFCC and C0

    • 3 MFCC, C0, and energy

• -e double

  • floor value of raw filter-bank output \((0 < \epsilon)\)

• infile str

  • double-type windowed sequence or spectrum

• stdout

  • double-type MFCCs

The below example extracts the 12-th order MFCCs from data.short. The analysis condition is that: frame length is 10 ms, frame shift is 25 ms, and sampling rate is 16 kHz. A pre-emphais filter and the hamming window are applied to the input signal.

x2x +sd data.short |
  frame -l 400 -p 160 -n 1 |
  dfs -b 1 -0.97 |
  window -l 400 -L 512 -w 1 -n 0 |
  mfcc -l 512 -n 40 -c 22 -m 12 -L 64 -H 4000 -o 1 |
  delta -m 12 -d -0.5 0.0 0.5 -d 0.25 0.0 -0.5 0.0 0.25 > data.mfcc

The corresponding HTK config file is shown as below.

SOURCEFORMAT = NOHEAD
SOURCEKIND   = WAVEFORM
SOURCERATE   = 625.0
TARGETKIND   = MFCC_E_D_A
TARGETRATE   = 100000.0
WINDOWSIZE   = 250000.0
USEHAMMING   = T
USEPOWER     = F
RAWENERGY    = F
ENORMALIZE   = F
PREEMCOEF    = 0.97
NUMCHANS     = 40
CEPLIFTER    = 22
NUMCEPS      = 12
LOFREQ       = 64
HIFREQ       = 4000
DELTAWINDOW  = 1
ACCWINDOW    = 1

The correspondence between the HTK’s configuration and the SPTK’s command option for extracting MFCC is shown in the following table.

Parameter	HTK	SPTK
Frame length	WINDOWSIZE = _ (unit is 100 ns)	frame -l _ (unit is point)
Frame shift	TARGETRATE = _ (unit is 100 ns)	frame -p _ (unit is point)
Sampling rate	SOURCERATE = _ (unit is 100 ns)	mfcc -s _ (unit is kHz)
Subtract mean	ZMEANSOURCE = T	frame -z
Pre-emphasis coefficient	PREEMCOEF = _ (windowed waveform)	dfs -b 1 -_ (entire waveform)
Window	USEHAMMING = T	window -w 1 -n 0
FFT length	N/A (auto. calculated)	mfcc -l _ (same as input length)
Number of fbank channels	NUMCHANS = _	mfcc -n _
Lowest frequency	LOFREQ = _	mfcc -L _
Highest frequency	HIFREQ = _	mfcc -H _
Floor value of fbank	N/A (fixed value: 1.0)	mfcc -e _ (default value: 1.0)
Order of cepstrum	NUMCEPS = _	mfcc -m _
Liftering coefficient	CEPLIFTER = _	mfcc -c _
Output energy	TARGETKIND = MFCC_E	mfcc -o 1
Output 0th coefficient	TARGETKIND = MFCC_0	mfcc -o 2
Use raw energy	RAWENERGY = T	N/A (do not use raw)
Normalize log energy	ENORMALIZE = T	N/A (do not normalize)
Delta window size	DELTAWINDOW = _	delta -d _
Accel window size	ACCWINDOW = _	delta -d * -d _

Parameters:

• argc – [in] Number of arguments.

• argv – [in] Argument vector.

Returns:

0 on success, 1 on failure.

See also

fbank plp

class MelFrequencyCepstralCoefficientsAnalysis

Perform mel-frequency cepstral coefficients (MFCC) analysis.

The input is the half part of power spectrum:

\[ \begin{array}{cccc} |X(0)|^2, & |X(1)|^2, & \ldots, & |X(N/2)|^2, \end{array} \]

where \(N\) is the FFT length. The outputs are the \(M\)-th order MFCCs with the zeroth cepstral parameter:

\[ \begin{array}{ccccc} c(0), & \bar{c}(1), & \bar{c}(2), & \ldots, & \bar{c}(M) \end{array} \]

and the log-signal energy \(E\).

The MFCCs are calculated from mel-filter-bank outputs \(\{F(j)\}_{j=1}^C\) using the discrete cosine transform

\[ c(m) = \sqrt{\frac{2}{C}} \sum_{j=1}^C F(j) \cos \left( \frac{\pi m}{C} \left(j - \frac{1}{2}\right) \right), \]

and the liftering

\[ \bar{c}(m) = \left( 1 + \frac{L}{2} \sin \frac{\pi m}{L} \right) c(m), \]

where \(L\) is the liftering parameter.

[1] S. Young et al., “The HTK book,” Cambridge University Engineering Department, 2006.

Public Functions

MelFrequencyCepstralCoefficientsAnalysis(int fft_length, int num_channel, int num_order, int liftering_coefficient, double sampling_rate, double lowest_frequency, double highest_frequency, double floor)

Parameters:

• fft_length – [in] Number of FFT bins, \(N\).

• num_channel – [in] Number of channels, \(C\).

• num_order – [in] Order of cepstral coefficients, \(M\).

• liftering_coefficient – [in] A parameter of liftering, \(L\).

• sampling_rate – [in] Sampling rate in Hz.

• lowest_frequency – [in] Lowest frequency in Hz.

• highest_frequency – [in] Highest frequency in Hz.

• floor – [in] Floor value of raw filter-bank output.

inline int GetFftLength() const

Returns:

FFT size.

inline int GetNumChannel() const

Returns:

Number of channels.

inline int GetNumOrder() const

Returns:

Order of cepstral coefficients.

inline int GetLifteringCoefficient() const

Returns:

Liftering coefficient.

inline bool IsValid() const

Returns:

True if this object is valid.

bool Run(const std::vector<double> &power_spectrum, std::vector<double> *mfcc, double *energy, MelFrequencyCepstralCoefficientsAnalysis::Buffer *buffer) const

Parameters:

• power_spectrum – [in] \((N/2+1)\)-length power spectrum.

• mfcc – [out] \(M\)-th order MFCCs.

• energy – [out] Signal energy \(E\) (optional).

• buffer – [out] Buffer.

Returns:

True on success, false on failure.

class Buffer

Buffer for MelFrequencyCepstralCoefficientsAnalysis class.