usages.doc

采用matlab进行语音处理的软件包

----------------------------------------------------------------------------
NAME
        feature - extract feature vectors from a speech signal

SYNOPSIS
	feature [options] speech_sample

DESCRIPTION
        This program calculate MFCC cepstral coefficients and their
derivative from a given signal. Three different feature sets or their
combinations can be specified.  An automatic starting/end point 
detection function is built-in which can remove the silent parts at
the beginning and end of the signal.

OPTIONS

	-o feature output file [stdout]
	-b [512] starting sample
	-w [256] analysis window size
	-s [128] window moving step
	-r [10] MFCC Order (before -M, G, I options)
	-T [1] Class membership
        -S *not* swap byte order, (SUN Big endian; PC, VAX little endian)
        -E Endpoint detection           [off]
	-M output MFCC			[off]
	-G output Delta MFCC		[off]
	-I output Accu  MFCC		[off]

EXAMPLES
        feature -o foo.pat -w 256 -s 128 -E -M -G -T 1 foo.wav

	extract 10 MFCC and 10 Delta MFCC coefficients from each 
frame of 256 samples. The class membership is "1" for this speech
token, vector patterns will be *appended* to the file "foo.pat".


----------------------------------------------------------------------------
NAME
	gensym - transform feature vector sequence to discrete 
observation symbols.

SYNOPSIS
	gensym [options] patternfile codebook

DESCRIPTION
	This program turns real vector sequences to discrete observation 
symbols which are used for training and test discrete hidden
Markov models. Internally, it invoke the LBG algorithm to general 
codebook if [-g] option is specified and find the nearest neighbor
for each input vector.

OPTIONS
	-o symbol file [stdout]
	-g generate codebook from the given data
	-n [32] number of code vectors, 2, 4, 8, 16,...
	-v verbose

EXAMPLES
        gensym -v -g -n 16 eset_tr.pat eset_tr.cod > eset_tr.seq
        gensym eset_te.pat eset_tr.cod > eset_te.seq


	The first example generates a codebook from the training 
patterns and encodes the training pattern, redirecting the sequence to
train.seq file. The second example uses the same codebook generated 
from the training  data and encodes the test patterns.


----------------------------------------------------------------------------
NAME
	dtwcmp - comparing test patterns with all training patterns by 
Dynamic time-wrapping (DTW) algorithm.

SYNOPSIS
	dtwcmp [options] reference_pattern test_pattern

DESCRIPTION
        This is the classical speech recognition method which align a test
speech token with each of reference patterns. 

OPTIONS
	 -d distance file name [none]
	 -k [1] k nearest neighbor
	 -v verbose

EXAMPLES
        dtwcmp -v -k 1 eset_tr.pat eset_te.pat

        This command find the nearest training pattern for each test pattern
based on DTW alignment and gives recognition results. The training patterns
"eset_tr.pat" and test patterns "eset_te.pat" are generated with the "feature"
utility from speech data.



----------------------------------------------------------------------------
NAME
        chmm - training program for continuous hidden Markov models with
Gaussian mixture density functions.

SYNOPSIS
	chmm [options] configfile patternfile modelfile

DESCRIPTION
	This is the state-of-the-art technique currently used for speech 
recognition. Each class of speech is modeled by a HMM with 
Gaussian mixture density functions and trained by the Baum-Welch 
algorithm. The model structure can be specified in the "configfile", 
but the most commonly used structure is the left-to-right model. In 
the configfile, "1" means the transition from this state (row) to 
another state (column) is allowed, otherwise, specify "0". 

OPTIONS
        -m [4] Number of Gaussian mixture components
	-s [1996] Random seed
	-i [0] Train specified model, 0 all models
	-v verbose

EXAMPLES
        chmm -v -v 8_state.cfg eset_tr.pat eset_tr.chm

        This command generates HMMs for each class in the "eset_tr.pat"
file and save the model parameters in "eset_tr.chm" for later test. Note,
most commands can give more than one "-v" option so that more 
messages will be dumped.


----------------------------------------------------------------------------
NAME
	viterbi - test program for hidden Markov models with Gaussian 
mixture density functions.

SYNOPSIS
	viterbi [options] patternfile modelfile

DESCRIPTION
	This is the companion program for the "chmm". It uses either 
viterbi decoding or forward scores to classify each pattern in the 
"patternfile".

OPTIONS
	-o [-] Output probability
	-p [-] Backtracking path
	-t Viterbi score, [Default forward variable]
	-v verbose

EXAMPLES
        viterbi -v eset_te.pat eset_tr.chm
        viterbi -p pathfile -t eset_te.pat eset_tr.chm

	The first command uses forward scores to test each pattern in 
"eset_te.pat" file and the second example use viterbi score. The align
paths are written in pathfile.



----------------------------------------------------------------------------
NAME
	dhmm - training program for discrete hidden Markov models.

SYNOPSIS
	dhmm [options] train_sequence modelfile

DESCRIPTION
	This program trains left-to-right discrete observation HMMs 
from given training sequences. The training sequences are generated 
with "gensym" program.


OPTIONS
	-s [6] : number of states
	-n [32] : number of symbols (codebook size)
	-i [0] : train single model, 0 all models
	-S [9373] : random seed
	-e [0.01] : stop training if prob. improving less than that
	-v verbose

EXAMPLES
        dhmm -v -v -n 16 eset_tr.seq eset_tr.dhm

	This command will generate discrete HMM from the training 
sequences. Note, the number of observation symbols should equal the 
codebook size used for generating sequences.


----------------------------------------------------------------------------
NAME
	dtest - test program for discrete hidden Markov models.

SYNOPSIS
	dtest [options] test_sequence modelfile

DESCRIPTION
	This is the companion program for the "dhmm". It uses forward 
scores to classify each sequence in the "test_sequence".

OPTIONS
	-i [none] Sequence ID (1:correct, 0:wrong)
	-p [none] write out prob. to a file
	-v verbose

EXAMPLES
        dtest -i eset_te.id eset_te.seq eset_tr.dhm

	The command gives the recognition result and tell you which 
sequence is correct in "eset_te.id" file ("1" correct, "0" wrong).