mfcc.h File Reference

Definitions for MFCC computation. More...

#include <sent/stddefs.h>
#include <ctype.h>

Include dependency graph for mfcc.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Defines
#define	DEF_SMPPERIOD   625
	Default sampling period in 100ns (625 = 16kHz).
#define	DEF_FRAMESIZE   400
	Default Window size in samples, similar to WINDOWSIZE in HTK (unit is different).
#define	DEF_FFTNUM   512
	Number of FFT steps.
#define	DEF_FRAMESHIFT   160
	Default frame shift length in samples.
#define	DEF_PREENPH   0.97
	Default pre-emphasis coefficient, corresponds to PREEMCOEF in HTK.
#define	DEF_MFCCDIM   12
	Default number of MFCC dimension, corresponds to NUMCEPS in HTK.
#define	DEF_CEPLIF   22
	Default cepstral Liftering coefficient, corresponds to CEPLIFTER in HTK.
#define	DEF_FBANK   24
	Default number of filterbank channels, corresponds to NUMCHANS in HTK.
#define	DEF_DELWIN   2
	Default delta window size, corresponds to DELTAWINDOW in HTK.
#define	DEF_ACCWIN   2
	Default acceleration window size, corresponds to ACCWINDOW in HTK.
#define	DEF_SILFLOOR   50.0
	Default energy silence floor in dBs, corresponds to SILFLOOR in HTK.
#define	DEF_ESCALE   0.1
	Default scaling coefficient of log energy, corresponds to ESCALE in HTK.
#define	DEF_SSALPHA   2.0
	Default alpha coefficient for spectral subtraction.
#define	DEF_SSFLOOR   0.5
	Default flooring coefficient for spectral subtraction.
Functions
void	make_costbl_hamming (int framesize)
void	make_fft_table (int n)
void	make_costbl_makemfcc (int fbank_num, int mfcc_dim)
void	make_sintbl_wcep (int lifter, int mfcc_dim)
void	Hamming (float *wave, int framesize)
void	PreEmphasise (float *wave, Value para)
float	Mel (int k, float fres)
FBankInfo	InitFBank (Value para)
void	FreeFBank (FBankInfo fb)
void	FFT (float *xRe, float *xIm, int p)
void	MakeFBank (float *wave, double *fbank, FBankInfo info, Value para, float *ssbuf)
void	MakeMFCC (double *fbank, float *mfcc, Value para)
float	CalcC0 (double *fbank, Value para)
float	CalcLogRawE (float *wave, int framesize)
void	ZMeanFrame (float *wave, int framesize)
void	WeightCepstrum (float *mfcc, Value para)
void	Delta (float **c, int frame, Value para)
void	Accel (float **c, int frame, Value para)
void	NormaliseLogE (float **c, int frame_num, Value para)
int	Wav2MFCC (SP16 *wave, float **mfcc, Value para, int nSamples, float *ssbuf, int ssbuflen)
void	CMN (float **mfcc, int frame_num, int dim)
void	WMP_init (Value para, float **bf, float *ssbuf, int ssbuflen)
void	WMP_calc (float *mfcc, float *bf, Value para, float *ssbuf)
DeltaBuf *	WMP_deltabuf_new (int veclen, int windowlen)
void	WMP_deltabuf_free (DeltaBuf *db)
void	WMP_deltabuf_prepare (DeltaBuf *db)
boolean	WMP_deltabuf_proceed (DeltaBuf *db, float *new_mfcc)
boolean	WMP_deltabuf_flush (DeltaBuf *db)
void	CMN_realtime_init (int dimension, float weight)
void	CMN_realtime_prepare ()
void	CMN_realtime (float *mfcc, int dim)
void	CMN_realtime_update ()
boolean	CMN_load_from_file (char *filename, int dim)
boolean	CMN_save_to_file (char *filename)
float *	new_SS_load_from_file (char *filename, int *slen)
float *	new_SS_calculate (SP16 *wave, int wavelen, Value para, int *slen)

---

Detailed Description

Definitions for MFCC computation.

Author:

Akinobu LEE

Date:

Fri Feb 11 03:40:52 2005

This file contains structures and default values for extracting speech parameter vectors of Mel-Frequency Cepstral Cefficients (MFCC). The default values here are the ones used in the standard acoustic models distributed together with Julius, and some of them have different value from HTK defaults. So be careful of the default values.

See also:

libsent/src/wav2mfcc/wav2mfcc.c

libsent/src/wav2mfcc/wav2mfcc-pipe.c

julius/wav2mfcc.c

julius/realtime-1stpass.c

Revision

1.4

Definition in file mfcc.h.

---

Function Documentation

void make_costbl_hamming (  int  framesize  )

Generate table for hamming window. Parameters: framesize  [in] window size Definition at line 74 of file wav2mfcc-buffer.c. Referenced by InitFBank(), and new_SS_calculate().

void make_fft_table (  int  n  )

Build tables for FFT. Parameters: n  [in] 2^n = FFT point Definition at line 105 of file wav2mfcc-buffer.c. Referenced by InitFBank(), and new_SS_calculate().

void make_costbl_makemfcc (  int  fbank_num, int  mfcc_dim )

Generate table for DCT operation to make mfcc from fbank. Parameters: fbank_num  [in] number of filer banks mfcc_dim  [in] number of dimensions in MFCC Definition at line 138 of file wav2mfcc-buffer.c. Referenced by InitFBank().

void make_sintbl_wcep (  int  lifter, int  mfcc_dim )

Generate table for weighing cepstrum. Parameters: lifter  [in] cepstral liftering coefficient mfcc_dim  [in] number of dimensions in MFCC Definition at line 176 of file wav2mfcc-buffer.c. Referenced by InitFBank().

void Hamming (  float *  wave, int  framesize )

Apply hamming window. Parameters: wave  [i/o] waveform data in the current frame framesize  [in] frame size Definition at line 354 of file wav2mfcc-buffer.c. Referenced by new_SS_calculate(), Wav2MFCC(), and WMP_calc().

void PreEmphasise (  float *  wave, Value  para )

Apply pre-emphasis filter. Parameters: wave  [i/o] waveform data in the current frame para  [in] configuration parameters Definition at line 374 of file wav2mfcc-buffer.c. Referenced by new_SS_calculate(), Wav2MFCC(), and WMP_calc().

float Mel (  int  k, float  fres )

Return mel-frequency. Parameters: k  [in] channel number of filter bank fres  [in] constant value computed by "1.0E7 / (para.smp_period * fb.fftN * 700.0)" Returns: the mel frequency. Definition at line 423 of file wav2mfcc-buffer.c. Referenced by InitFBank().

FBankInfo InitFBank (  Value  para  )

Build filterbank information and generate tables for MFCC comptutation. Parameters: para  [in] configuration parameters Returns: the generated filterbank information. Definition at line 435 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_init().

void FreeFBank (  FBankInfo  fb  )

Free FBankInfo. Parameters: fb  [in] filterbank information Definition at line 529 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC().

void FFT (  float *  xRe, float *  xIm, int  p )

Apply FFT Parameters: xRe  [i/o] real part of waveform xIm  [i/o] imaginal part of waveform p  [in] 2^p = FFT point Definition at line 782 of file wav2mfcc-buffer.c. Referenced by MakeFBank(), and new_SS_calculate().

void MakeFBank (  float *  wave, double *  fbank, FBankInfo  fb, Value  para, float *  ssbuf )

Convert wave -> (spectral subtraction) -> mel-frequency filterbank Parameters: wave  [in] waveform data in the current frame fbank  [out] the resulting mel-frequency filterbank fb  [in] filterbank information para  [in] configuration parameters ssbuf  [in] noise spectrum, or NULL if not apply subtraction Definition at line 547 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_calc().

void MakeMFCC (  double *  fbank, float *  mfcc, Value  para )

Apply DCT to filterbank to make MFCC. Parameters: fbank  [in] filterbank mfcc  [out] output MFCC vector para  [in] configuration parameters Definition at line 607 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_calc().

float CalcC0 (  double *  fbank, Value  para )

Calculate 0'th cepstral coefficient. Parameters: fbank  [in] filterbank para  [in] configuration parameters Returns: Definition at line 644 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_calc().

float CalcLogRawE (  float *  wave, int  framesize )

Calculate Log Raw Energy. Parameters: wave  [in] waveform data in the current frame framesize  [in] frame size Returns: the calculated log raw energy. Definition at line 316 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_calc().

void ZMeanFrame (  float *  wave, int  framesize )

Remove DC offset per frame Parameters: wave  [i/o] waveform data in the current frame framesize  [in] frame size Definition at line 337 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_calc().

void WeightCepstrum (  float *  mfcc, Value  para )

Re-scale cepstral coefficients. Parameters: mfcc  [i/o] a MFCC vector para  [in] configuration parameters Definition at line 389 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC(), and WMP_calc().

void Delta (  float **  c, int  frame, Value  para )

Calculate delta coefficients Parameters: c  [i/o] MFCC vectors, in which the delta coeff. will be appended. frame  [in] number of frames para  [in] configuration parameters Definition at line 695 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC().

void Accel (  float **  c, int  frame, Value  para )

Calculate acceleration coefficients. Parameters: c  [i/o] MFCC vectors, in which the delta coeff. will be appended. frame  [in] number of frames para  [in] configuration parameters Definition at line 744 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC().

void NormaliseLogE (  float **  mfcc, int  frame_num, Value  para )

Normalise log energy Parameters: mfcc  [i/o] array of MFCC vectors frame_num  [in] number of frames para  [in] configuration parameters Definition at line 663 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC().

int Wav2MFCC (  SP16 *  wave, float **  mfcc, Value  para, int  nSamples, float *  ssbuf, int  ssbuflen )

Convert wave data to MFCC. Also does spectral subtraction if ssbuf specified. Parameters: wave  [in] waveform data mfcc  [out] buffer to store the resulting MFCC parameter vector [t][0..veclen-1], should be already allocated para  [in] configuration parameters nSamples  [in] length of waveform data ssbuf  [in] buffer that holds noise spectrum to be subtracted from input, or NULL if not use spectral subtraction ssbuflen  [in] length of above, ignored when ssbuf is NULL Returns: the number of processed frames. Definition at line 217 of file wav2mfcc-buffer.c. Referenced by new_wav2mfcc().

void CMN (  float **  mfcc, int  frame_num, int  dim )

Cepstrum Mean Normalization (buffered) Cepstral mean will be computed within the given MFCC vectors. Parameters: mfcc  [i/o] array of MFCC vectors frame_num  [in] number of frames dim  [in] total dimension of MFCC vectors Definition at line 835 of file wav2mfcc-buffer.c. Referenced by Wav2MFCC().

void WMP_init (  Value  para, float **  bf, float *  ssbuf, int  ssbuflen )

initialize and setup buffers for a MFCC computataion. Parameters: para  [in] configuration parameters bf  [out] pointer to the entry point of workspace for FFT ssbuf  [in] noise spectrum, or NULL if not using spectral subtraction ssbuflen  [in] length of above, ignoredwhen ssbuf is NULL Definition at line 60 of file wav2mfcc-pipe.c. Referenced by RealTimeInit().

void WMP_calc (  float *  mfcc, float *  bf, Value  para, float *  ssbuf )

Calculate MFCC and log energy for one frame. Perform spectral subtraction if ssbuf is specified. Parameters: mfcc  [out] buffer to hold the resulting MFCC vector bf  [i/o] work area for FFT para  [in] configuration parameters ssbuf  [in] noise spectrum, or NULL if not using spectral subtraction Definition at line 95 of file wav2mfcc-pipe.c. Referenced by RealTimePipeLine().

DeltaBuf* WMP_deltabuf_new (  int  veclen, int  windowlen )

Allocate a new delta cycle buffer. Parameters: veclen  [in] length of a vector windowlen  [in] window width for computing delta Returns: pointer to newly allocated delta cycle buffer structure. Definition at line 141 of file wav2mfcc-pipe.c. Referenced by RealTimeInit().

void WMP_deltabuf_free (  DeltaBuf *  db  )

Destroy the delta cycle buffer. Parameters: db  [i/o] delta cycle buffer Definition at line 168 of file wav2mfcc-pipe.c.

void WMP_deltabuf_prepare (  DeltaBuf *  db  )

Reset and clear the delta cycle buffer. Parameters: db  [i/o] delta cycle buffer Definition at line 186 of file wav2mfcc-pipe.c. Referenced by RealTimePipeLinePrepare().

boolean WMP_deltabuf_proceed (  DeltaBuf *  db, float *  new_mfcc )

Store the given MFCC vector into the delta cycle buffer, and compute the latest delta coefficients. Parameters: db  [i/o] delta cycle buffer new_mfcc  [in] MFCC vector Returns: TRUE if next delta coeff. computed, in that case it is saved in db->delta[], or FALSE if delta is not yet computed by short of data. Definition at line 245 of file wav2mfcc-pipe.c. Referenced by RealTimeParam(), and RealTimePipeLine().

boolean WMP_deltabuf_flush (  DeltaBuf *  db  )

Flush the delta cycle buffer the delta coefficients left in the cycle buffer. Parameters: db  [i/o] delta cycle buffer Returns: TRUE if next delta coeff. computed, in that case it is saved in db->delta[], or FALSE if all delta computation has been flushed and no data is available. Definition at line 287 of file wav2mfcc-pipe.c. Referenced by RealTimeParam().

void CMN_realtime_init (  int  dimension, float  weight )

Initialize MAP-CMN at startup. Parameters: dimension  [in] vector dimension weight  [in] initial cepstral mean weight Definition at line 348 of file wav2mfcc-pipe.c. Referenced by RealTimeInit().

void CMN_realtime_prepare (   )

Prepare for MAP-CMN at start of each input Definition at line 375 of file wav2mfcc-pipe.c. Referenced by RealTimePipeLinePrepare().

void CMN_realtime (  float *  mfcc, int  dim )

Perform MAP-CMN for incoming MFCC vectors Parameters: mfcc  [in] MFCC vector dim  [in] dimension Definition at line 390 of file wav2mfcc-pipe.c. Referenced by RealTimeParam(), and RealTimePipeLine().

void CMN_realtime_update (   )

Update initial cepstral mean from previous utterances for next input. Definition at line 418 of file wav2mfcc-pipe.c. Referenced by RealTimeCMNUpdate().

boolean CMN_load_from_file (  char *  filename, int  dim )

Load CMN parameter from file. If the number of MFCC dimension in the file does not match the specified one, an error will occur. Parameters: filename  [in] file name dim  [in] required number of MFCC dimensions Returns: TRUE on success, FALSE on failure. Definition at line 518 of file wav2mfcc-pipe.c. Referenced by RealTimeInit().

boolean CMN_save_to_file (  char *  filename  )

Save the current CMN vector to a file. Parameters: filename  [in] filename to save the data. Returns: TRUE on success, FALSE on failure. Definition at line 562 of file wav2mfcc-pipe.c. Referenced by RealTimeCMNUpdate().

float* new_SS_load_from_file (  char *  filename, int *  slen )

Load a noise spectrum from file. Parameters: filename  [in] path name of noise spectrum file slen  [out] length of the returned buffer Returns: a newly allocated buffer that holds the loaded noise spectrum. Definition at line 65 of file ss.c. Referenced by new_wav2mfcc(), and RealTimeInit().

float* new_SS_calculate (  SP16 *  wave, int  wavelen, Value  para, int *  slen )

Compute average spectrum of audio input. This is used to estimate a noise spectrum from input samples. Parameters: wave  [in] input audio data sequence wavelen  [in] length of above para  [in] parameter slen  [out] length of returned buffer Returns: a newly allocated buffer that contains the calculated spectrum. Definition at line 109 of file ss.c. Referenced by new_wav2mfcc().

---