MFCC計算のための定義 [詳細]

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

データ構造
struct	Value
	mfcc configuration parameter values [詳細]
struct	FBankInfo
	Workspace for filterbank analysis. [詳細]
struct	DeltaBuf
	Cycle buffer for delta computation. [詳細]
struct	MFCCWork
	Work area for MFCC computation. [詳細]
struct	CMEAN
	Structure to hold sentence sum of MFCC for realtime CMN. [詳細]
struct	CMNWork
	Work area for real-time CMN. [詳細]
struct	ENERGYWork
	work area for energy normalization on real time input [詳細]
マクロ定義
#define	CPMAX 500
	DEBUG: define if you want to enable debug messages for sin/cos table operation.
#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.
関数
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.

説明

MFCC計算のための定義

このファイルには，音声波形データからMFCC形式の特徴量ベクトル系列を計算するための構造体の定義およびデフォルト値が含まれています．デフォルト値は Julius とともに配布されている音響モデルで使用している値であり，HTKのデフォルトとは値が異なる部分がありますので注意して下さい．

参照:: libsent/src/wav2mfcc/wav2mfcc.c; libsent/src/wav2mfcc/wav2mfcc-pipe.c; julius/wav2mfcc.c; julius/realtime-1stpass.c

作者:: Akinobu LEE

日付:: Fri Feb 11 03:40:52 2005

Revision:: 1.5

mfcc.h で定義されています。

マクロ定義

#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

mfcc.h の 47 行で定義されています。

参照元 CMN_realtime_update().

関数

MFCCWork* WMP_work_new ( Value * para )

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

引数:

para	[in] configuration parameters

戻り値:: pointer to the newly allocated work area.

mfcc-core.c の 620 行で定義されています。

参照元 j_final_fusion(), と 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.

引数:

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

mfcc-core.c の 658 行で定義されています。

参照元 RealTimeMFCC(), と Wav2MFCC().

void WMP_free ( MFCCWork * w )

Free all work area for MFCC computation.

引数:

w	[i/o] MFCC calculation work area

mfcc-core.c の 700 行で定義されています。

参照元 j_mfcccalc_free().

boolean InitFBank	(	MFCCWork *	w,
		Value *	para
	)

Build filterbank information and generate tables for MFCC comptutation.

引数:

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

戻り値:: the generated filterbank information.

mfcc-core.c の 229 行で定義されています。

参照元 WMP_work_new().

void FreeFBank ( FBankInfo * fb )

Free FBankInfo.

引数:

fb	[in] filterbank information

mfcc-core.c の 311 行で定義されています。

参照元 WMP_free().

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

Apply hamming window.

引数:

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

mfcc-core.c の 383 行で定義されています。

参照元 new_SS_calculate(), と WMP_calc().

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

Apply pre-emphasis filter.

引数:

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

mfcc-core.c の 367 行で定義されています。

参照元 new_SS_calculate(), と WMP_calc().

float Mel	(	int	k,
		float	fres
	)

Return mel-frequency.

引数:

k	[in] channel number of filter bank
fres	[in] constant value computed by "1.0E7 / (para.smp_period * fb.fftN * 700.0)"

戻り値:: the mel frequency.

mfcc-core.c の 161 行で定義されています。

参照元 InitFBank().

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

Apply FFT.

引数:

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

mfcc-core.c の 405 行で定義されています。

参照元 MakeFBank(), と new_SS_calculate().

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

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

引数:

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

mfcc-core.c の 458 行で定義されています。

参照元 WMP_calc().

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

Apply DCT to filterbank to make MFCC.

引数:

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

mfcc-core.c の 547 行で定義されています。

参照元 WMP_calc().

float CalcC0	(	MFCCWork *	w,
		Value *	para
	)

Calculate 0'th cepstral coefficient.

引数:

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

戻り値:

mfcc-core.c の 529 行で定義されています。

参照元 WMP_calc().

float CalcLogRawE	(	float *	wave,
		int	framesize
	)

Calculate Log Raw Energy.

引数:

wave	[in] waveform data in the current frame
framesize	[in] frame size

戻り値:: the calculated log raw energy.

mfcc-core.c の 347 行で定義されています。

参照元 WMP_calc().

void ZMeanFrame	(	float *	wave,
		int	framesize
	)

Remove DC offset per frame.

引数:

wave	[i/o] waveform data in the current frame
framesize	[in] frame size

mfcc-core.c の 328 行で定義されています。

参照元 new_SS_calculate(), と WMP_calc().

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

Re-scale cepstral coefficients.

引数:

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

mfcc-core.c の 582 行で定義されています。

参照元 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.

引数:

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

戻り値:: the number of processed frames.

wav2mfcc-buffer.c の 57 行で定義されています。

参照元 wav2mfcc().

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

Calculate delta coefficients.

引数:

c	[i/o] MFCC vectors, in which the delta coeff. will be appended.
frame	[in] number of frames
para	[in] configuration parameters

wav2mfcc-buffer.c の 142 行で定義されています。

参照元 Wav2MFCC().

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

Calculate acceleration coefficients.

引数:

c	[i/o] MFCC vectors, in which the delta coeff. will be appended.
frame	[in] number of frames
para	[in] configuration parameters

wav2mfcc-buffer.c の 180 行で定義されています。

参照元 Wav2MFCC().

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

Normalise log energy.

引数:

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

wav2mfcc-buffer.c の 110 行で定義されています。

参照元 Wav2MFCC().

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

Cepstrum Mean Normalization (buffered) Cepstral mean will be computed within the given MFCC vectors.

引数:

mfcc	[i/o] array of MFCC vectors
frame_num	[in] number of frames
dim	[in] total dimension of MFCC vectors

wav2mfcc-buffer.c の 219 行で定義されています。

参照元 Wav2MFCC().

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

Cepstrum Mean/Variance Normalization (buffered)

引数:

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

wav2mfcc-buffer.c の 248 行で定義されています。

参照元 Wav2MFCC().

DeltaBuf* WMP_deltabuf_new	(	int	veclen,
		int	windowlen
	)

Allocate a new delta cycle buffer.

引数:

veclen	[in] length of a vector
windowlen	[in] window width for computing delta

戻り値:: pointer to newly allocated delta cycle buffer structure.

wav2mfcc-pipe.c の 60 行で定義されています。

参照元 RealTimeInit().

void WMP_deltabuf_free ( DeltaBuf * db )

Destroy the delta cycle buffer.

引数:

db	[i/o] delta cycle buffer

wav2mfcc-pipe.c の 87 行で定義されています。

参照元 j_mfcccalc_free().

void WMP_deltabuf_prepare ( DeltaBuf * db )

Reset and clear the delta cycle buffer.

引数:

db	[i/o] delta cycle buffer

wav2mfcc-pipe.c の 105 行で定義されています。

参照元 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.

引数:

db	[i/o] delta cycle buffer
new_mfcc	[in] MFCC vector

戻り値:: 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.

wav2mfcc-pipe.c の 164 行で定義されています。

参照元 RealTimeMFCC(), と RealTimeParam().

boolean WMP_deltabuf_flush ( DeltaBuf * db )

Flush the delta cycle buffer the delta coefficients left in the cycle buffer.

引数:

db	[i/o] delta cycle buffer

戻り値:: 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.

wav2mfcc-pipe.c の 206 行で定義されています。

参照元 RealTimeParam().

CMNWork* CMN_realtime_new	(	Value *	para,
		float	weight
	)

Initialize MAP-CMN at startup.

引数:

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

wav2mfcc-pipe.c の 247 行で定義されています。

参照元 RealTimeInit().

void CMN_realtime_free ( CMNWork * c )

Free work area for real-time CMN.

引数:

c	[i/o] CMN calculation work area

wav2mfcc-pipe.c の 285 行で定義されています。

参照元 j_mfcccalc_free().

void CMN_realtime_prepare ( CMNWork * c )

Prepare for MAP-CMN at start of each input.

引数:

c	[i/o] CMN calculation work area

wav2mfcc-pipe.c の 309 行で定義されています。

参照元 init_param(), と RealTimeResume().

void CMN_realtime	(	CMNWork *	c,
		float *	mfcc
	)

Perform MAP-CMN for incoming MFCC vectors.

引数:

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

wav2mfcc-pipe.c の 328 行で定義されています。

参照元 RealTimeMFCC(), と RealTimeParam().

void CMN_realtime_update	(	CMNWork *	c,
		HTK_Param *	param
	)

Update initial cepstral mean from previous utterances for next input.

引数:

c	[i/o] CMN calculation work area

wav2mfcc-pipe.c の 389 行で定義されています。

参照元 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.

引数:

c	[i/o] CMN calculation work area
filename	[in] file name

戻り値:: TRUE on success, FALSE on failure.

wav2mfcc-pipe.c の 522 行で定義されています。

参照元 RealTimeInit().

boolean CMN_save_to_file	(	CMNWork *	c,
		char *	filename
	)

Save the current CMN vector to a file.

引数:

c	[i/o] CMN calculation work area
filename	[in] filename to save the data.

戻り値:: TRUE on success, FALSE on failure.

wav2mfcc-pipe.c の 579 行で定義されています。

参照元 RealTimeCMNUpdate().

void energy_max_init ( ENERGYWork * energy )

Initialize work area for energy normalization on live input.

This should be called once on startup.

引数:

energy [in] energy normalization work area

wav2mfcc-pipe.c の 633 行で定義されています。

参照元 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.

引数:

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

wav2mfcc-pipe.c の 646 行で定義されています。

参照元 reset_mfcc().

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

Peform energy normalization using maximum of last input.

引数:

energy	[in] energy normalization work area
f	[in] raw energy
para	[in] MFCC computation configuration parameter

戻り値:: value of the normalized log energy.

wav2mfcc-pipe.c の 663 行で定義されています。

参照元 RealTimeMFCC().

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

Load a noise spectrum from file.

引数:

filename	[in] path name of noise spectrum file
slen	[out] length of the returned buffer

戻り値:: a newly allocated buffer that holds the loaded noise spectrum.

ss.c の 66 行で定義されています。

参照元 RealTimeInit(), と 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.

引数:

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

戻り値:: a newly allocated buffer that contains the calculated spectrum.

ss.c の 111 行で定義されています。

参照元 wav2mfcc().

void undef_para ( Value * para )

Reset configuration parameters for MFCC computation.

引数:

para	[out] feature extraction parameters

para.c の 39 行で定義されています。

参照元 initialize_GSHMM(), と jconf_set_default_values_am().

void make_default_para ( Value * para )

Set Julius default parameters for MFCC computation.

引数:

para	[out] feature extraction parameters

para.c の 85 行で定義されています。

参照元 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.

引数:

para	[out] feature extraction parameters

para.c の 118 行で定義されています。

参照元 jconf_set_default_values_am().

void apply_para	(	Value *	dst,
		Value *	src
	)

Merge two configuration parameters for MFCC computation.

引数:

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

para.c の 145 行で定義されています。

参照元 initialize_GMM(), と initialize_HMM().

boolean htk_config_file_parse	(	char *	HTKconffile,
		Value *	para
	)

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

引数:

HTKconffile	[in] HTK Config file path name
para	[out] MFCC parameter to set

戻り値:: TRUE on success, FALSE on failure.

para.c の 193 行で定義されています。

参照元 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.

引数:

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

para.c の 320 行で定義されています。

参照元 initialize_GMM(), と initialize_HMM().

void put_para	(	FILE *	fp,
		Value *	para
	)

Output acoustic analysis configuration parameters to stdout.

引数:

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

para.c の 356 行で定義されています。

参照元 print_engine_info().

データ構造

マクロ定義

関数

説明

マクロ定義

関数