As mentioned in section 5.1, the tool HLIST provides a dual rôle in HTK. Firstly, it can be used for examining the contents of speech data files. In general, HLIST displays three types of information
As an example, suppose that the file called timit.wav holds speech waveform data using the TIMIT format. The command
HList -h -e 49 -F TIMIT timit.wavwould display the source header information and the first 50 samples of the file. The output would look something like the following
----------------------------- Source: timit.wav ---------------------------
Sample Bytes: 2 Sample Kind: WAVEFORM
Num Comps: 1 Sample Period: 62.5 us
Num Samples: 31437 File Format: TIMIT
------------------------------ Samples: 0->49 -----------------------------
0: 8 -4 -1 0 -2 -1 -3 -2 0 0
10: -1 0 -1 -2 -1 1 0 -1 -2 1
20: -2 0 0 0 2 1 -2 2 1 0
30: 1 0 0 -1 4 2 0 -1 4 0
40: 2 2 1 -1 -1 1 1 2 1 1
------------------------------------ END ----------------------------------
The source information confirms that the file contains WAVEFORM
data with 2 byte samples and 31437 samples in total. The sample period is
which corresponds to a 16kHz sample rate.
The displayed data is numerically small because it corresponds to leading silence.
Any part of the file could be viewed by suitable choice of the begin and end
sample indices. For example,
HList -s 5000 -e 5049 -F TIMIT timit.wavwould display samples 5000 through to 5049. The output might look like the following
---------------------------- Samples: 5000->5049 -------------------------- 5000: 85 -116 -159 -252 23 99 69 92 79 -166 5010: -100 -123 -111 48 -19 15 111 41 -126 -304 5020: -189 91 162 255 80 -134 -174 -55 57 155 5030: 90 -1 33 154 68 -149 -70 91 165 240 5040: 297 50 13 72 187 189 193 244 198 128 ------------------------------------ END ----------------------------------
The second use of HLIST is to check that input conversions are being performed properly. Suppose that the above TIMIT format file is part of a database to be used for training a recogniser and that mel-frequency cepstra are to be used along with energy and the first differential coefficients. Suitable configuration parameters needed to achieve this might be as follows
# Wave -> MFCC config file
SOURCEFORMAT = TIMIT # same as -F TIMIT
TARGETKIND = MFCC_E_D # MFCC + Energy + Deltas
TARGETRATE = 100000 # 10ms frame rate
WINDOWSIZE = 200000 # 20ms window
NUMCHANS = 24 # num filterbank chans
NUMCEPS = 8 # compute c1 to c8
HLIST can be used to check this. For example, typing
HList -C config -o -h -t -s 100 -e 104 -i 9 timit.wavwill cause the waveform file to be converted, then the source header, the target header and parameter vectors 100 through to 104 to be listed. A typical output would be as follows
------------------------------ Source: timit.wav ---------------------------
Sample Bytes: 2 Sample Kind: WAVEFORM
Num Comps: 1 Sample Period: 62.5 us
Num Samples: 31437 File Format: TIMIT
------------------------------------ Target --------------------------------
Sample Bytes: 72 Sample Kind: MFCC_E_D
Num Comps: 18 Sample Period: 10000.0 us
Num Samples: 195 File Format: HTK
-------------------------- Observation Structure ---------------------------
x: MFCC-1 MFCC-2 MFCC-3 MFCC-4 MFCC-5 MFCC-6 MFCC-7 MFCC-8 E
Del-1 Del-2 Del-3 Del-4 Del-5 Del-6 Del-7 Del-8 DelE
------------------------------ Samples: 100->104 ---------------------------
100: 3.573 -19.729 -1.256 -6.646 -8.293 -15.601 -23.404 10.988 0.834
3.161 -1.913 0.573 -0.069 -4.935 2.309 -5.336 2.460 0.080
101: 3.372 -16.278 -4.683 -3.600 -11.030 -8.481 -21.210 10.472 0.777
0.608 -1.850 -0.903 -0.665 -2.603 -0.194 -2.331 2.180 0.069
102: 2.823 -15.624 -5.367 -4.450 -12.045 -15.939 -22.082 14.794 0.830
-0.051 0.633 -0.881 -0.067 -1.281 -0.410 1.312 1.021 0.005
103: 3.752 -17.135 -5.656 -6.114 -12.336 -15.115 -17.091 11.640 0.825
-0.002 -0.204 0.015 -0.525 -1.237 -1.039 1.515 1.007 0.015
104: 3.127 -16.135 -5.176 -5.727 -14.044 -14.333 -18.905 15.506 0.833
-0.034 -0.247 0.103 -0.223 -1.575 0.513 1.507 0.754 0.006
------------------------------------- END ----------------------------------
The target header information shows that the converted data consists of 195 parameter vectors, each vector having 18 components and being 72 bytes in size. The structure of each parameter vector is displayed as a simple sequence of floating-point numbers. The layout information described in section 5.7 can be used to interpret the data. However, including the -o option, as in the example, causes HLIST to output a schematic of the observation structure. Thus, it can be seen that the first row of each sample contains the static coefficients and the second contains the delta coefficients. The energy is in the final column. The command line option -i 9 controls the number of values displayed per line and can be used to aid in the visual interpretation of the data. Notice finally that the command line option -F TIMIT was not required in this case because the source format was specified in the configuration file.
It should be stressed that when HLIST displays parameterised data, it does so in exactly the form that observations are passed to a HTK tool. So, for example, if the above data was input to a system built using 3 data streams, then this can be simulated by using the command line option -n to set the number of streams. For example, typing
HList -C config -n 3 -o -s 100 -e 101 -i 9 timit.wavwould result in the following output
------------------------ Observation Structure ----------------------- nTotal=18 nStatic=8 nDel=16 eSep=T x.1: MFCC-1 MFCC-2 MFCC-3 MFCC-4 MFCC-5 MFCC-6 MFCC-7 MFCC-8 x.2: Del-1 Del-2 Del-3 Del-4 Del-5 Del-6 Del-7 Del-8 x.3: E DelE -------------------------- Samples: 100->101 ------------------------- 100.1: 3.573 -19.729 -1.256 -6.646 -8.293 -15.601 -23.404 10.988 100.2: 3.161 -1.913 0.573 -0.069 -4.935 2.309 -5.336 2.460 100.3: 0.834 0.080 101.1: 3.372 -16.278 -4.683 -3.600 -11.030 -8.481 -21.210 10.472 101.2: 0.608 -1.850 -0.903 -0.665 -2.603 -0.194 -2.331 2.180 101.3: 0.777 0.069 --------------------------------- END --------------------------------Notice that the data is identical to the previous case, but it has been re-organised into separate streams.