tomwalters@54
|
1 #!/bin/bash
|
tomwalters@54
|
2 # Copyright 2010 Thomas Walters <tom@acousticscale.org>
|
tomwalters@54
|
3 #
|
tomwalters@54
|
4 # Run a series of experiments which compare MFCC features generated by HTK to
|
tomwalters@54
|
5 # AIM features generated using AIM-C using a series of syllable recogntiton
|
tomwalters@54
|
6 # tasks.
|
tomwalters@54
|
7 # This script expects the HTK binaries and AIM-C AIMCopy binary to be present
|
tomwalters@54
|
8 # in the PATH.
|
tomwalters@54
|
9
|
tomwalters@54
|
10 # Set these to be the location of your input database, and desired output
|
tomwalters@54
|
11 # locations.
|
tomwalters@54
|
12 SYLLABLES_DATABASE_TAR=/media/sounds/cnbh-syllables.tar
|
tomwalters@54
|
13 SOUNDS_ROOT=/mnt/experiments/sounds/
|
tomwalters@54
|
14 FEATURES_ROOT=/mnt/experiments/features/
|
tomwalters@54
|
15 HMMS_ROOT=/mnt/experiments/hmms/
|
tomwalters@54
|
16
|
tomwalters@54
|
17 # Number of cores on the experimental machine. Various scripts will try to use
|
tomwalters@54
|
18 # this if it's set.
|
tomwalters@54
|
19 NUMBER_OF_CORES=2
|
tomwalters@54
|
20
|
tomwalters@54
|
21 # Fail if any command fails
|
tomwalters@54
|
22 set -e
|
tomwalters@54
|
23
|
tomwalters@54
|
24 # Fail if any variable is unset
|
tomwalters@54
|
25 set -u
|
tomwalters@54
|
26
|
tomwalters@54
|
27 if [ ! -d $SOUNDS_ROOT ]; then
|
tomwalters@58
|
28 sudo mkdir -p $SOUNDS_ROOT
|
tomwalters@58
|
29 sudo chown `whoami` $SOUNDS_ROOT
|
tomwalters@54
|
30 fi
|
tomwalters@54
|
31
|
tomwalters@54
|
32 # Untar the CNBH syllables database, and convert the files from FLAC to WAV
|
tomwalters@54
|
33 if [ ! -e $SOUNDS_ROOT/.untar_db_success ]; then
|
tomwalters@54
|
34 tar -x -C $SOUNDS_ROOT -f $SYLLABLES_DATABASE_TAR
|
tomwalters@54
|
35 touch $SOUNDS_ROOT/.untar_db_success
|
tomwalters@54
|
36 fi
|
tomwalters@54
|
37
|
tomwalters@54
|
38 # Convert the database to .WAV format and place it in $SOUNDS_ROOT/clean
|
tomwalters@60
|
39 ./cnbh-syllables/feature_generation/convert_flac_to_wav.sh $SOUNDS_ROOT
|
tomwalters@54
|
40
|
tomwalters@54
|
41 # Generate versions of the CNBH syllables spoke pattern with a range of
|
tomwalters@54
|
42 # signal-to-noise ratios (SNRs). The versions are put in the directory
|
tomwalters@54
|
43 # ${SOUNDS_ROOT}/${SNR}_dB/ for each SNR in $SNRS.
|
tomwalters@54
|
44 SNRS="30 27 24 21 18 15 12 9 6 3 0"
|
tomwalters@59
|
45 ./cnbh-syllables/feature_generation/pink_noise.sh $SOUNDS_ROOT/clean/ $SNRS
|
tomwalters@54
|
46
|
tomwalters@54
|
47 # Make the list of all feature drectories
|
tomwalters@54
|
48 FEATURE_DIRS="clean"
|
tomwalters@54
|
49 for SNR in $SNRS; do
|
tomwalters@54
|
50 FEATURE_DIRS="$FEATURE_DIRS snr_${SNR}dB"
|
tomwalters@54
|
51 done
|
tomwalters@54
|
52
|
tomwalters@54
|
53 # Generate feature sets (for the full range of SNRs in $FEATURE_DIRS)
|
tomwalters@54
|
54 # 1. Standard MFCC features
|
tomwalters@54
|
55 # 2. AIM features
|
tomwalters@54
|
56 # 3. MFCC features with optimal VTLN
|
tomwalters@58
|
57
|
tomwalters@58
|
58
|
tomwalters@58
|
59 if [ ! -d $FEATURES_ROOT ]; then
|
tomwalters@58
|
60 sudo mkdir -p $FEATURES_ROOT
|
tomwalters@58
|
61 sudo chown `whoami` $FEATURES_ROOT
|
tomwalters@58
|
62 fi
|
tomwalters@58
|
63
|
tomwalters@54
|
64 for SOURCE_SNR in $FEATURE_DIRS; do
|
tomwalters@54
|
65
|
tomwalters@65
|
66 if [ ! -e $FEATURES_ROOT/mfcc/$SOURCE_SNR/.make_mfcc_features_success]; then
|
tomwalters@54
|
67 mkdir -p $FEATURES_ROOT/mfcc/$SOURCE_SNR/
|
tomwalters@54
|
68 # Generate the list of files to convert
|
tomwalters@54
|
69 ./cnbh-syllables/feature_generation/gen_hcopy_aimcopy_script.sh $FEATURES_ROOT/mfcc/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/
|
tomwalters@54
|
70 # Run the conversion
|
tomwalters@54
|
71 ./cnbh-syllables/feature_generation/run_hcopy.sh $FEATURES_ROOT/mfcc/$SOURCE_SNR/ $NUMBER_OF_CORES
|
tomwalters@54
|
72 touch $FEATURES_ROOT/mfcc/$SOURCE_SNR/.make_mfcc_features_success
|
tomwalters@65
|
73 fi
|
tomwalters@54
|
74
|
tomwalters@65
|
75 if [ ! -e $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/.make_mfcc_vtln_features_success]; then
|
tomwalters@54
|
76 mkdir -p $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/
|
tomwalters@54
|
77 # Generate the file list and run the conversion (all one step, since this
|
tomwalters@54
|
78 # version uses a different configuraiton for each talker)
|
tomwalters@54
|
79 ./cnbh-syllables/feature_generation/run_mfcc_vtln_conversion.sh $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/
|
tomwalters@54
|
80 touch $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/.make_mfcc_vtln_features_success
|
tomwalters@65
|
81 fi
|
tomwalters@54
|
82
|
tomwalters@65
|
83 if [ ! -e $FEATURES_ROOT/aim/$SOURCE_SNR/.make_aim_features_success]; then
|
tomwalters@54
|
84 mkdir -p $FEATURES_ROOT/aim/$SOURCE_SNR/
|
tomwalters@54
|
85 ./cnbh-syllables/feature_generation/gen_hcopy_aimcopy_script.sh $FEATURES_ROOT/aim/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/
|
tomwalters@54
|
86 # Run the conversion
|
tomwalters@54
|
87 ./cnbh-syllables/feature_generation/run_aimcopy.sh $FEATURES_ROOT/aim/$SOURCE_SNR/ $NUMBER_OF_CORES
|
tomwalters@54
|
88 touch $FEATURES_ROOT/aim/$SOURCE_SNR/.make_aim_features_success
|
tomwalters@65
|
89 fi
|
tomwalters@54
|
90 done
|
tomwalters@54
|
91
|
tomwalters@54
|
92 # Now run a bunch of experiments.
|
tomwalters@54
|
93 # For each of the feature types, we want to run HMMs with a bunch of
|
tomwalters@54
|
94 # parameters.
|
tomwalters@54
|
95 TRAINING_ITERATIONS="0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20"
|
tomwalters@54
|
96 TESTING_ITERATIONS="1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20"
|
tomwalters@54
|
97 HMM_STATES="3 4 5 6 7 8"
|
tomwalters@54
|
98 HMM_OUTPUT_COMPONENTS=""
|
tomwalters@54
|
99
|
tomwalters@54
|
100
|