1SPHINX_FE(1) General Commands Manual SPHINX_FE(1)
2
6 sphinx_fe - Convert audio files to acoustic feature files
7
9 sphinx_fe [ options ]...
10
12 This program converts audio files (in either Microsoft WAV, NIST
13 Sphere, or raw format) to acoustic feature files for input to batch-
14 mode speech recognition. The resulting files are also useful for vari‐
15 ous other things. A list of options follows:
16 -alpha Preemphasis parameter
18 -argfile
20 file (e.g. feat.params from an acoustic model) to read parame‐
21 ters from. This will override anything set in other command
22 line arguments.
23 -blocksize
25 Number of samples to read at a time.
26 -build_outdirs
28 Create missing subdirectories in output directory
29 -c file for batch processing
31 -cep2spec
33 Input is cepstral files, output is log spectral files
34 -di directory, input file names are relative to this, if defined
36 -dither
38 Add 1/2-bit noise
39 -do directory, output files are relative to this
41 -doublebw
43 Use double bandwidth filters (same center freq)
44 -ei extension to be applied to all input files
46 -eo extension to be applied to all output files
48 -example
50 Shows example of how to use the tool
51 -frate Frame rate
53 -help Shows the usage of the tool
55 -i audio input file
57 -input_endian
59 Endianness of input data, big or little, ignored if NIST or MS
60 Wav
61 -lifter
63 Length of sin-curve for liftering, or 0 for no liftering.
64 -logspec
66 Write out logspectral files instead of cepstra
67 -lowerf
69 Lower edge of filters
70 -mach_endian
72 Endianness of machine, big or little
73 -mswav Defines input format as Microsoft Wav (RIFF)
75 -ncep Number of cep coefficients
77 -nchans
79 Number of channels of data (interlaced samples assumed)
80 -nfft Size of FFT
82 -nfilt Number of filter banks
84 -nist Defines input format as NIST sphere
86 -npart Number of parts to run in (supersedes -nskip and -runlen if non-
88 zero)
89 -nskip If a control file was specified, the number of utterances to
91 skip at the head of the file
92 -o cepstral output file
94 -ofmt Format of output files - one of sphinx, htk, text.
96 -part Index of the part to run (supersedes -nskip and -runlen if non-
98 zero)
99 -raw Defines input format as raw binary data
101 -remove_dc
103 Remove DC offset from each frame
104 -remove_noise
106 Remove noise with spectral subtraction in mel-energies
107 -remove_silence
109 Enables VAD, removes silence frames from processing
110 -round_filters
112 Round mel filter frequencies to DFT points
113 -runlen
115 If a control file was specified, the number of utterances to
116 process, or -1 for all
117 -samprate
119 Sampling rate
120 -seed Seed for random number generator; if less than zero, pick our
122 own
123 -smoothspec
125 Write out cepstral-smoothed logspectral files
126 -spec2cep
128 Input is log spectral files, output is cepstral files
129 -sph2pipe
131 Input is NIST sphere (possibly with Shorten), use sph2pipe to
132 convert
133 -transform
135 Which type of transform to use to calculate cepstra (legacy,
136 dct, or htk)
137 -unit_area
139 Normalize mel filters to unit area
140 -upperf
142 Upper edge of filters
143 -vad_postspeech
145 Num of silence frames to keep after from speech to silence.
146 -vad_prespeech
148 Num of speech frames to keep before silence to speech.
149 -vad_startspeech
151 Num of speech frames to trigger vad from silence to speech.
152 -vad_threshold
154 Threshold for decision between noise and silence frames. Log-
155 ratio between signal level and noise level.
156 -verbose
158 Show input filenames
159 -warp_params
161 defining the warping function
162 -warp_type
164 Warping function type (or shape)
165 -whichchan
167 Channel to process (numbered from 1), or 0 to mix all channels
168 -wlen Hamming window length
170 Currently the only kind of features supported are MFCCs (mel-frequency
172 cepstral coefficients). There are numerous options which control the
173 properties of the output features. It is VERY important that you docu‐
174 ment the specific set of flags used to create any given set of feature
175 files, since this information is NOT recorded in the files themselves,
176 and any mismatch between the parameters used to extract features for
177 recognition and those used to extract features for training will cause
178 recognition to fail.
179
181 Written by numerous people at CMU from 1994 onwards. This manual page
182 by David Huggins-Daines <dhuggins@cs.cmu.edu>
183
185 Copyright © 1994-2007 Carnegie Mellon University. See the file COPYING
186 included with this package for more information.
187 2007-08-27 SPHINX_FE(1)