libsent/include/sent/mfcc.h File Reference

Definitions for MFCC computation. More...

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

Go to the source code of this file.

Data Structures

struct  Value
 mfcc configuration parameter values More...
struct  FBankInfo
 Workspace for filterbank analysis. More...
struct  DeltaBuf
 Cycle buffer for delta computation. More...
struct  MFCCWork
 Work area for MFCC computation. More...
struct  CMEAN
 Structure to hold sentence sum of MFCC for realtime CMN. More...
struct  CMNWork
 Work area for real-time CMN. More...
struct  ENERGYWork
 work area for energy normalization on real time input More...

Defines

#define CPMAX   500
 DEBUG: define if you want to enable debug messages for sin/cos table operation Maximum number of frames to store ceptral mean for realtime CMN update.
#define CPSTEP   5
 allocate step of cmean list per sentence
#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   1.0
 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.
#define VALUE_VERSION   3
 Integer version number of Value, for embedding.

Functions

MFCCWork * WMP_work_new (Value *para)
 Setup work area for parameters, values, buffers, tables to compute MFCC vectors, with a given parameter configurations.
void WMP_calc (MFCCWork *w, float *mfcc, Value *para)
 Calculate MFCC and log energy for one frame.
void WMP_free (MFCCWork *w)
 Free all work area for MFCC computation.
boolean InitFBank (MFCCWork *w, Value *para)
 Build filterbank information and generate tables for MFCC comptutation.
void FreeFBank (FBankInfo *fb)
 Free FBankInfo.
void Hamming (float *wave, int framesize, MFCCWork *w)
 Apply hamming window.
void PreEmphasise (float *wave, int framesize, float preEmph)
 Apply pre-emphasis filter.
float Mel (int k, float fres)
 Return mel-frequency.
void FFT (float *xRe, float *xIm, int p, MFCCWork *w)
 Apply FFT.
void MakeFBank (float *wave, MFCCWork *w, Value *para)
 Convert wave -> (spectral subtraction) -> mel-frequency filterbank.
void MakeMFCC (float *mfcc, Value *para, MFCCWork *w)
 Apply DCT to filterbank to make MFCC.
float CalcC0 (MFCCWork *w, Value *para)
 Calculate 0'th cepstral coefficient.
float CalcLogRawE (float *wave, int framesize)
 Calculate Log Raw Energy.
void ZMeanFrame (float *wave, int framesize)
 Remove DC offset per frame.
void WeightCepstrum (float *mfcc, Value *para, MFCCWork *w)
 Re-scale cepstral coefficients.
int Wav2MFCC (SP16 *wave, float **mfcc, Value *para, int nSamples, MFCCWork *w)
 Convert wave data to MFCC.
void Delta (float **c, int frame, Value *para)
 Calculate delta coefficients.
void Accel (float **c, int frame, Value *para)
 Calculate acceleration coefficients.
void NormaliseLogE (float **c, int frame_num, Value *para)
 Normalise log energy.
void CMN (float **mfcc, int frame_num, int dim)
 Cepstrum Mean Normalization (buffered) Cepstral mean will be computed within the given MFCC vectors.
void MVN (float **mfcc, int frame_num, Value *para)
 Cepstrum Mean/Variance Normalization (buffered).
DeltaBuf * WMP_deltabuf_new (int veclen, int windowlen)
 Allocate a new delta cycle buffer.
void WMP_deltabuf_free (DeltaBuf *db)
 Destroy the delta cycle buffer.
void WMP_deltabuf_prepare (DeltaBuf *db)
 Reset and clear the delta cycle buffer.
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.
boolean WMP_deltabuf_flush (DeltaBuf *db)
 Flush the delta cycle buffer the delta coefficients left in the cycle buffer.
CMNWork * CMN_realtime_new (Value *para, float weight)
 Initialize MAP-CMN at startup.
void CMN_realtime_free (CMNWork *c)
 Free work area for real-time CMN.
void CMN_realtime_prepare (CMNWork *c)
 Prepare for MAP-CMN at start of each input.
void CMN_realtime (CMNWork *c, float *mfcc)
 Perform MAP-CMN for incoming MFCC vectors.
void CMN_realtime_update (CMNWork *c, HTK_Param *param)
 Update initial cepstral mean from previous utterances for next input.
boolean CMN_load_from_file (CMNWork *c, char *filename)
 Load CMN parameter from file.
boolean CMN_save_to_file (CMNWork *c, char *filename)
 Save the current CMN vector to a file.
void energy_max_init (ENERGYWork *energy)
 Initialize work area for energy normalization on live input.
void energy_max_prepare (ENERGYWork *energy, Value *para)
 Prepare values for energy normalization on live input.
LOGPROB energy_max_normalize (ENERGYWork *energy, LOGPROB f, Value *para)
 Peform energy normalization using maximum of last input.
float * new_SS_load_from_file (char *filename, int *slen)
 Load a noise spectrum from file.
float * new_SS_calculate (SP16 *wave, int wavelen, int *slen, MFCCWork *w, Value *para)
 Compute average spectrum of audio input.
void undef_para (Value *para)
 Reset configuration parameters for MFCC computation.
void make_default_para (Value *para)
 Set Julius default parameters for MFCC computation.
void make_default_para_htk (Value *para)
 Set HTK default configuration parameters for MFCC computation.
void apply_para (Value *dst, Value *src)
 Merge two configuration parameters for MFCC computation.
boolean htk_config_file_parse (char *HTKconffile, Value *para)
 Read and parse an HTK Config file, and set the specified option values.
void calc_para_from_header (Value *para, short param_type, short vec_size)
 Set acoustic analysis parameters from HTK HMM definition header information.
void put_para (FILE *fp, Value *para)
 Output acoustic analysis configuration parameters to stdout.

Detailed Description

Definitions for MFCC computation.

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

Author:
Akinobu LEE
Date:
Fri Feb 11 03:40:52 2005
Revision
1.4

Definition in file mfcc.h.

Function Documentation

MFCCWork* WMP_work_new ( Value *  para  ) 

Setup work area for parameters, values, buffers, tables to compute MFCC vectors, with a given parameter configurations.

Parameters:
para [in] configuration parameters
Returns:
pointer to the newly allocated work area.

Definition at line 614 of file mfcc-core.c.

Referenced by j_final_fusion(), and j_mfcccalc_new().

void WMP_calc ( MFCCWork *  w,
float *  mfcc,
Value *  para 
)

Calculate MFCC and log energy for one frame.

Perform spectral subtraction if ssbuf is specified.

Parameters:
w [i/o] MFCC calculation work area
mfcc [out] buffer to hold the resulting MFCC vector
para [in] configuration parameters

Definition at line 652 of file mfcc-core.c.

Referenced by RealTimeMFCC(), and Wav2MFCC().

void WMP_free ( MFCCWork *  w  ) 

Free all work area for MFCC computation.

Parameters:
w [i/o] MFCC calculation work area

Definition at line 694 of file mfcc-core.c.

Referenced by j_mfcccalc_free().

boolean InitFBank ( MFCCWork *  w,
Value *  para 
)

Build filterbank information and generate tables for MFCC comptutation.

Parameters:
w [i/o] MFCC calculation work area
para [in] configuration parameters
Returns:
the generated filterbank information.

Definition at line 223 of file mfcc-core.c.

Referenced by WMP_work_new().

void FreeFBank ( FBankInfo *  fb  ) 

Free FBankInfo.

Parameters:
fb [in] filterbank information

Definition at line 305 of file mfcc-core.c.

Referenced by WMP_free().

void Hamming ( float *  wave,
int  framesize,
MFCCWork *  w 
)

Apply hamming window.

Parameters:
wave [i/o] waveform data in the current frame
framesize [in] frame size
w [i/o] MFCC calculation work area

Definition at line 377 of file mfcc-core.c.

Referenced by new_SS_calculate(), and WMP_calc().

void PreEmphasise ( float *  wave,
int  framesize,
float  preEmph 
)

Apply pre-emphasis filter.

Parameters:
wave [i/o] waveform data in the current frame
framesize [i/o] frame size in samples
preEmph [in] pre-emphasis coef.

Definition at line 361 of file mfcc-core.c.

Referenced by new_SS_calculate(), 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 155 of file mfcc-core.c.

Referenced by InitFBank().

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

Apply FFT.

Parameters:
xRe [i/o] real part of waveform
xIm [i/o] imaginal part of waveform
p [in] 2^p = FFT point
w [i/o] MFCC calculation work area

Definition at line 399 of file mfcc-core.c.

void MakeFBank ( float *  wave,
MFCCWork *  w,
Value *  para 
)

Convert wave -> (spectral subtraction) -> mel-frequency filterbank.

Parameters:
wave [in] waveform data in the current frame
w [i/o] MFCC calculation work area
para [in] configuration parameters

Definition at line 452 of file mfcc-core.c.

Referenced by WMP_calc().

void MakeMFCC ( float *  mfcc,
Value *  para,
MFCCWork *  w 
)

Apply DCT to filterbank to make MFCC.

Parameters:
mfcc [out] output MFCC vector
para [in] configuration parameters
w [i/o] MFCC calculation work area

Definition at line 541 of file mfcc-core.c.

Referenced by WMP_calc().

float CalcC0 ( MFCCWork *  w,
Value *  para 
)

Calculate 0'th cepstral coefficient.

Parameters:
w [i/o] MFCC calculation work area
para [in] configuration parameters
Returns:

Definition at line 523 of file mfcc-core.c.

Referenced by 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 341 of file mfcc-core.c.

Referenced by 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 322 of file mfcc-core.c.

Referenced by new_SS_calculate(), and WMP_calc().

void WeightCepstrum ( float *  mfcc,
Value *  para,
MFCCWork *  w 
)

Re-scale cepstral coefficients.

Parameters:
mfcc [i/o] a MFCC vector
para [in] configuration parameters
w [i/o] MFCC calculation work area

Definition at line 576 of file mfcc-core.c.

Referenced by WMP_calc().

int Wav2MFCC ( SP16 *  wave,
float **  mfcc,
Value *  para,
int  nSamples,
MFCCWork *  w 
)

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
w [i/o] MFCC calculation work area
Returns:
the number of processed frames.

Definition at line 57 of file wav2mfcc-buffer.c.

Referenced by wav2mfcc().

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 142 of file wav2mfcc-buffer.c.

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 180 of file wav2mfcc-buffer.c.

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 110 of file wav2mfcc-buffer.c.

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 219 of file wav2mfcc-buffer.c.

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

Cepstrum Mean/Variance Normalization (buffered).

Parameters:
mfcc [i/o] array of MFCC vectors
frame_num [in] number of frames
para [in] configuration parameters

Definition at line 248 of file wav2mfcc-buffer.c.

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 60 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 87 of file wav2mfcc-pipe.c.

Referenced by j_mfcccalc_free().

void WMP_deltabuf_prepare ( DeltaBuf *  db  ) 

Reset and clear the delta cycle buffer.

Parameters:
db [i/o] delta cycle buffer

Definition at line 105 of file wav2mfcc-pipe.c.

Referenced by reset_mfcc().

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 164 of file wav2mfcc-pipe.c.

Referenced by RealTimeMFCC(), and RealTimeParam().

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 206 of file wav2mfcc-pipe.c.

Referenced by RealTimeParam().

CMNWork* CMN_realtime_new ( Value *  para,
float  weight 
)

Initialize MAP-CMN at startup.

Parameters:
para [in] MFCC computation configuration parameter
weight [in] initial cepstral mean weight

Definition at line 247 of file wav2mfcc-pipe.c.

Referenced by RealTimeInit().

void CMN_realtime_free ( CMNWork *  c  ) 

Free work area for real-time CMN.

Parameters:
c [i/o] CMN calculation work area

Definition at line 285 of file wav2mfcc-pipe.c.

Referenced by j_mfcccalc_free().

void CMN_realtime_prepare ( CMNWork *  c  ) 

Prepare for MAP-CMN at start of each input.

Parameters:
c [i/o] CMN calculation work area

Definition at line 309 of file wav2mfcc-pipe.c.

Referenced by init_param(), and RealTimeResume().

void CMN_realtime ( CMNWork *  c,
float *  mfcc 
)

Perform MAP-CMN for incoming MFCC vectors.

Parameters:
c [i/o] CMN calculation work area
mfcc [in] MFCC vector

Definition at line 328 of file wav2mfcc-pipe.c.

Referenced by RealTimeMFCC(), and RealTimeParam().

void CMN_realtime_update ( CMNWork *  c,
HTK_Param *  param 
)

Update initial cepstral mean from previous utterances for next input.

Parameters:
c [i/o] CMN calculation work area

Definition at line 389 of file wav2mfcc-pipe.c.

Referenced by RealTimeCMNUpdate().

boolean CMN_load_from_file ( CMNWork *  c,
char *  filename 
)

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:
c [i/o] CMN calculation work area
filename [in] file name
Returns:
TRUE on success, FALSE on failure.

Definition at line 522 of file wav2mfcc-pipe.c.

Referenced by RealTimeInit().

boolean CMN_save_to_file ( CMNWork *  c,
char *  filename 
)

Save the current CMN vector to a file.

Parameters:
c [i/o] CMN calculation work area
filename [in] filename to save the data.
Returns:
TRUE on success, FALSE on failure.

Definition at line 579 of file wav2mfcc-pipe.c.

Referenced by RealTimeCMNUpdate().

void energy_max_init ( ENERGYWork *  energy  ) 

Initialize work area for energy normalization on live input.

This should be called once on startup.

Parameters:
energy [in] energy normalization work area

Definition at line 629 of file wav2mfcc-pipe.c.

Referenced by RealTimeInit().

void energy_max_prepare ( ENERGYWork *  energy,
Value *  para 
)

Prepare values for energy normalization on live input.

This should be called before each input segment.

Parameters:
energy [in] energy normalization work area
para [in] MFCC computation configuration parameter

Definition at line 642 of file wav2mfcc-pipe.c.

Referenced by reset_mfcc().

LOGPROB energy_max_normalize ( ENERGYWork *  energy,
LOGPROB  f,
Value *  para 
)

Peform energy normalization using maximum of last input.

Parameters:
energy [in] energy normalization work area
f [in] raw energy
para [in] MFCC computation configuration parameter
Returns:
value of the normalized log energy.

Definition at line 659 of file wav2mfcc-pipe.c.

Referenced by RealTimeMFCC().

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 66 of file ss.c.

Referenced by RealTimeInit(), and wav2mfcc().

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

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
slen [out] length of returned buffer
w [i/o] MFCC calculation work area
para [in] parameter
Returns:
a newly allocated buffer that contains the calculated spectrum.

Definition at line 111 of file ss.c.

Referenced by wav2mfcc().

void undef_para ( Value *  para  ) 

Reset configuration parameters for MFCC computation.

Parameters:
para [out] feature extraction parameters

Definition at line 39 of file para.c.

Referenced by initialize_GSHMM(), and jconf_set_default_values_am().

void make_default_para ( Value *  para  ) 

Set Julius default parameters for MFCC computation.

Parameters:
para [out] feature extraction parameters

Definition at line 85 of file para.c.

Referenced by jconf_set_default_values_am().

void make_default_para_htk ( Value *  para  ) 

Set HTK default configuration parameters for MFCC computation.

This will be refered when parameters are given as HTK Config file.

Parameters:
para [out] feature extraction parameters

Definition at line 118 of file para.c.

Referenced by jconf_set_default_values_am().

void apply_para ( Value *  dst,
Value *  src 
)

Merge two configuration parameters for MFCC computation.

Parameters:
dst [out] feature extraction parameters to set to
src [out] feature extraction parameters to set from

Definition at line 145 of file para.c.

Referenced by initialize_GMM(), and initialize_HMM().

boolean htk_config_file_parse ( char *  HTKconffile,
Value *  para 
)

Read and parse an HTK Config file, and set the specified option values.

Parameters:
HTKconffile [in] HTK Config file path name
para [out] MFCC parameter to set
Returns:
TRUE on success, FALSE on failure.

Definition at line 193 of file para.c.

Referenced by opt_parse().

void calc_para_from_header ( Value *  para,
short  param_type,
short  vec_size 
)

Set acoustic analysis parameters from HTK HMM definition header information.

Parameters:
para [out] acoustic analysis parameters
param_type [in] parameter type specified at HMM header
vec_size [in] vector size type specified at HMM header

Definition at line 320 of file para.c.

Referenced by initialize_GMM(), and initialize_HMM().

void put_para ( FILE *  fp,
Value *  para 
)

Output acoustic analysis configuration parameters to stdout.

Parameters:
fp [in] file pointer
para [in] configuration parameter

Definition at line 356 of file para.c.

Referenced by print_engine_info().

Generated on Thu Jul 23 12:14:19 2009 for Julius by  doxygen 1.5.1