tomwalters@54: #!/bin/bash
tomwalters@54: # Copyright 2010 Thomas Walters <tom@acousticscale.org>
tomwalters@54: #
tomwalters@54: # Run a series of experiments which compare MFCC features generated by HTK to
tomwalters@54: # AIM features generated using AIM-C using a series of syllable recogntiton
tomwalters@54: # tasks.
tomwalters@90: # This script expects to be run from within the AIM-C source tree.
tomwalters@101: # It builds the HTK binaries and AIM-C AIMCopy binary if they're not
tomwalters@101: # present.
tomwalters@90: # The following environment varaibles should be set before this script is run:
tomwalters@90: # SYLLABLES_DATABASE_URL - URL of a tar file containing the CNBH syllables
tomwalters@90: # database in FLAC format
tomwalters@101: # HTK_USERNAME and HTK_PASSWORD - username and password for the site at
tomwalters@90: # http://htk.eng.cam.ac.uk/
tomwalters@98: # NUMBER_OF_CORES - total number of machine cores
tomwalters@90: 
tomwalters@54: # Set these to be the location of your input database, and desired output
tomwalters@101: # locations. (Note: the user running this script needs write permissions on
tomwalters@101: # the $WORKING_VOLUME.)
tomwalters@101: WORKING_VOLUME=/mnt/scratch1
tomwalters@101: 
tomwalters@101: SYLLABLES_DATABASE_TAR=$WORKING_VOLUME/001-downloaded_sounds_data/cnbh-syllables.tar
tomwalters@101: SOUNDS_ROOT=$WORKING_VOLUME/002-sounds/
tomwalters@101: FEATURES_ROOT=$WORKING_VOLUME/003-features/
tomwalters@101: HMMS_ROOT=$WORKING_VOLUME/004-hmms/
tomwalters@101: HTK_ROOT=$WORKING_VOLUME/software/htk/
tomwalters@101: AIMC_ROOT=$WORKING_VOLUME/software/aimc/
tomwalters@54: 
tomwalters@54: # Number of cores on the experimental machine. Various scripts will try to use
tomwalters@54: # this if it's set.
tomwalters@98: # NUMBER_OF_CORES=8
tomwalters@54: 
tomwalters@54: # Fail if any command fails
tomwalters@54: set -e
tomwalters@54: 
tomwalters@54: # Fail if any variable is unset
tomwalters@54: set -u
tomwalters@54: 
tomwalters@101: ######
tomwalters@101: # Step 001 - Get the sounds database
tomwalters@67: if [ ! -e $SYLLABLES_DATABASE_TAR ]; then
tomwalters@101:   mkdir -p `dirname $SYLLABLES_DATABASE_TAR`
tomwalters@67:   wget -O $SYLLABLES_DATABASE_TAR $SYLLABLES_DATABASE_URL
tomwalters@67: fi
tomwalters@67: 
tomwalters@54: if [ ! -d $SOUNDS_ROOT ]; then
tomwalters@101:   mkdir -p $SOUNDS_ROOT
tomwalters@54: fi
tomwalters@54: 
tomwalters@101: # Untar the CNBH syllables database, and convert the files from FLAC to WAV.
tomwalters@54: if [ ! -e $SOUNDS_ROOT/.untar_db_success ]; then
tomwalters@54:   tar -x -C $SOUNDS_ROOT -f $SYLLABLES_DATABASE_TAR
tomwalters@54:   touch $SOUNDS_ROOT/.untar_db_success
tomwalters@54: fi
tomwalters@54: 
tomwalters@54: # Convert the database to .WAV format and place it in $SOUNDS_ROOT/clean
tomwalters@74: echo "Converting CNBH-syllables database from FLAC to WAV..."
tomwalters@60: ./cnbh-syllables/feature_generation/convert_flac_to_wav.sh $SOUNDS_ROOT
tomwalters@101: #
tomwalters@101: ######
tomwalters@54: 
tomwalters@101: #####
tomwalters@101: # Step 002 -
tomwalters@54: # Generate versions of the CNBH syllables spoke pattern with a range of
tomwalters@54: # signal-to-noise ratios (SNRs). The versions are put in the directory
tomwalters@54: # ${SOUNDS_ROOT}/${SNR}_dB/ for each SNR in $SNRS.
tomwalters@96: SNRS="30 27 24 21 18 15 12 9 6 3 0"
tomwalters@96: #SNRS="30" # For testing
tomwalters@75: ./cnbh-syllables/feature_generation/pink_noise.sh $SOUNDS_ROOT/clean/ "$SNRS"
tomwalters@54: 
tomwalters@54: # Make the list of all feature drectories
tomwalters@54: FEATURE_DIRS="clean"
tomwalters@54: for SNR in $SNRS; do
tomwalters@54:   FEATURE_DIRS="$FEATURE_DIRS snr_${SNR}dB"
tomwalters@54: done
tomwalters@54: 
tomwalters@54: # Generate feature sets (for the full range of SNRs in $FEATURE_DIRS)
tomwalters@54: # 1. Standard MFCC features
tomwalters@54: # 2. AIM features
tomwalters@101: # 3. MFCC features with optimal VTLN
tomwalters@58: 
tomwalters@58: if [ ! -d $FEATURES_ROOT ]; then
tomwalters@101:  mkdir -p $FEATURES_ROOT
tomwalters@58: fi
tomwalters@58: 
tomwalters@101: if [ ! -e $HTK_ROOT/.htk_installed_success ]; then
tomwalters@101:   ./HTK/install_htk.sh $HTK_ROOT
tomwalters@70: fi
tomwalters@69: 
tomwalters@101: if [ ! -e $AIMC_ROOT/.aimc_build_success ]; then
tomwalters@101:  ./aimc/build_aimc.sh $AIMC_ROOT
tomwalters@72: fi
tomwalters@72: 
tomwalters@54: for SOURCE_SNR in $FEATURE_DIRS; do
tomwalters@66:   if [ ! -e $FEATURES_ROOT/mfcc/$SOURCE_SNR/.make_mfcc_features_success ]; then
tomwalters@54:     mkdir -p $FEATURES_ROOT/mfcc/$SOURCE_SNR/
tomwalters@54:     # Generate the list of files to convert
tomwalters@78:     ./cnbh-syllables/feature_generation/gen_hcopy_aimcopy_script.sh $FEATURES_ROOT/mfcc/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/ htk
tomwalters@54:     # Run the conversion
tomwalters@101:     ./cnbh-syllables/feature_generation/run_hcopy.sh $FEATURES_ROOT/mfcc/$SOURCE_SNR/ $NUMBER_OF_CORES
tomwalters@101:     touch $FEATURES_ROOT/mfcc/$SOURCE_SNR/.make_mfcc_features_success
tomwalters@65:   fi
tomwalters@54: 
tomwalters@66:   if [ ! -e $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/.make_mfcc_vtln_features_success ]; then
tomwalters@54:     mkdir -p $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/
tomwalters@54:     # Generate the file list and run the conversion (all one step, since this
tomwalters@75:     # version uses a different configuration for each talker)
tomwalters@101:     ./cnbh-syllables/feature_generation/run_mfcc_vtln_conversion.sh $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/
tomwalters@101:     touch $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/.make_mfcc_vtln_features_success
tomwalters@65:   fi
tomwalters@54: 
tomwalters@66:   if [ ! -e $FEATURES_ROOT/aim/$SOURCE_SNR/.make_aim_features_success ]; then
tomwalters@101:     mkdir -p $FEATURES_ROOT/aim/$SOURCE_SNR/
tomwalters@78:     ./cnbh-syllables/feature_generation/gen_hcopy_aimcopy_script.sh $FEATURES_ROOT/aim/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/ ""
tomwalters@54:     # Run the conversion
tomwalters@98:     ./cnbh-syllables/feature_generation/run_aimcopy.sh $FEATURES_ROOT/aim/$SOURCE_SNR/ $NUMBER_OF_CORES
tomwalters@98:     touch $FEATURES_ROOT/aim/$SOURCE_SNR/.make_aim_features_success
tomwalters@65:   fi
tomwalters@101: done
tomwalters@54: 
tomwalters@101: mkdir -p $HMMS_ROOT
tomwalters@82: 
tomwalters@54: # Now run a bunch of experiments.
tomwalters@54: # For each of the feature types, we want to run HMMs with a bunch of
tomwalters@54: # parameters.
tomwalters@101: TRAINING_ITERATIONS="0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15" # 16 17 18 19 20"
tomwalters@101: #TESTING_ITERATIONS="0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20"
tomwalters@101: TESTING_ITERATIONS="15"
tomwalters@101: #HMM_STATES="3 4 5 6 7 8"
tomwalters@101: HMM_STATES="4"
tomwalters@101: #HMM_OUTPUT_COMPONENTS="1 2 3 4 5 6 7"
tomwalters@101: HMM_OUTPUT_COMPONENTS="4"
tomwalters@91: 
tomwalters@93: run_train_test () {
tomwalters@96: # TODO(tom): Make sure that the training SNR is generated first
tomwalters@93: for SOURCE_SNR in $FEATURE_DIRS; do
tomwalters@93: WORK=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/$SOURCE_SNR/$TALKERS/
tomwalters@93: mkdir -p $WORK
tomwalters@93: FEATURES_DIR=$FEATURES_ROOT/$FEATURE_CLASS/$SOURCE_SNR/
tomwalters@104: SPOKE_PATTERN_FILE=`pwd`/cnbh-syllables/run_training_and_testing/train_test_sets/gen_spoke_points/spoke_pattern.txt
tomwalters@93: 
tomwalters@93: ./cnbh-syllables/run_training_and_testing/train_test_sets/generate_train_test_lists.sh \
tomwalters@93:     $TALKERS \
tomwalters@93:     $WORK \
tomwalters@93:     $FEATURES_DIR \
tomwalters@93:     $FEATURE_SUFFIX
tomwalters@93: 
tomwalters@96: TRAINING_SCRIPT=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/$TRAINING_SNR/$TALKERS/training_script
tomwalters@96: TRAINING_MASTER_LABEL_FILE=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/$TRAINING_SNR/$TALKERS/training_master_label_file
tomwalters@96: 
tomwalters@93: TESTING_SCRIPT=$WORK/testing_script
tomwalters@96: TESTING_MASTER_LABEL_FILE=$WORK/testing_master_label_file
tomwalters@93: 
tomwalters@93: ./cnbh-syllables/run_training_and_testing/gen_htk_base_files.sh $WORK
tomwalters@93: 
tomwalters@93: ./cnbh-syllables/run_training_and_testing/test_features.sh \
tomwalters@93:     "$WORK" \
tomwalters@93:     "$FEATURES_ROOT/$FEATURE_CLASS/$SOURCE_SNR/" \
tomwalters@93:     "$FEATURE_SUFFIX" \
tomwalters@93:     "$HMM_STATES" \
tomwalters@93:     "$HMM_OUTPUT_COMPONENTS" \
tomwalters@93:     "$TRAINING_ITERATIONS" \
tomwalters@93:     "$TESTING_ITERATIONS" \
tomwalters@93:     "$FEATURE_SIZE" \
tomwalters@93:     "$FEATURE_TYPE" \
tomwalters@93:     "$TRAINING_SCRIPT" \
tomwalters@93:     "$TESTING_SCRIPT" \
tomwalters@95:     "$TRAINING_MASTER_LABEL_FILE" \
tomwalters@104:     "$TESTING_MASTER_LABEL_FILE" \
tomwalters@104:     "$SPOKE_PATTERN_FILE"
tomwalters@93: done
tomwalters@93: }
tomwalters@93: 
tomwalters@91: ########################
tomwalters@91: # Standard MFCCs
tomwalters@75: FEATURE_CLASS=mfcc
tomwalters@79: FEATURE_SUFFIX=htk
tomwalters@75: FEATURE_SIZE=39
tomwalters@75: FEATURE_TYPE=MFCC_0_D_A
tomwalters@91: TALKERS=inner_talkers
tomwalters@96: TRAINING_SNR=clean
tomwalters@91: run_train_test
tomwalters@91: ########################
tomwalters@54: 
tomwalters@91: ########################
tomwalters@91: # Standard MFCCs
tomwalters@91: # Train on extrema
tomwalters@91: FEATURE_CLASS=mfcc
tomwalters@91: FEATURE_SUFFIX=htk
tomwalters@91: FEATURE_SIZE=39
tomwalters@91: FEATURE_TYPE=MFCC_0_D_A
tomwalters@91: TALKERS=outer_talkers
tomwalters@96: TRAINING_SNR=clean
tomwalters@91: run_train_test
tomwalters@91: ########################
tomwalters@91: 
tomwalters@91: ########################
tomwalters@91: # MFCCs with VTLN
tomwalters@91: FEATURE_CLASS=mfcc_vtln
tomwalters@91: FEATURE_SUFFIX=htk
tomwalters@91: FEATURE_SIZE=39
tomwalters@91: FEATURE_TYPE=MFCC_0_D_A
tomwalters@78: TALKERS=inner_talkers
tomwalters@96: TRAINING_SNR=clean
tomwalters@91: run_train_test
tomwalters@91: ########################
tomwalters@91: 
tomwalters@91: ########################
tomwalters@91: # MFCCs with VTLN
tomwalters@91: # Train on extrema
tomwalters@91: FEATURE_CLASS=mfcc_vtln
tomwalters@91: FEATURE_SUFFIX=htk
tomwalters@91: FEATURE_SIZE=39
tomwalters@91: FEATURE_TYPE=MFCC_0_D_A
tomwalters@91: TALKERS=outer_talkers
tomwalters@96: TRAINING_SNR=clean
tomwalters@91: run_train_test
tomwalters@91: ########################
tomwalters@91: 
tomwalters@101: AIM_FEATURE_SUFFIXES="slice_1_no_cutoff ssi_profile_no_cutoff slice_1_cutoff ssi_profile_cutoff smooth_nap_profile"
tomwalters@101: for f in $AIM_FEATURE_SUFFIXES
tomwalters@101: do
tomwalters@91: ########################
tomwalters@91: # AIM Features
tomwalters@101: # Inner talkers
tomwalters@101: FEATURE_CLASS=aim
tomwalters@101: FEATURE_SUFFIX=$f
tomwalters@101: FEATURE_SIZE=12
tomwalters@101: FEATURE_TYPE=USER_E_D_A
tomwalters@101: TALKERS=inner_talkers
tomwalters@101: TRAINING_SNR=clean
tomwalters@101: run_train_test
tomwalters@91: ########################
tomwalters@91: 
tomwalters@101: ########################
tomwalters@101: # AIM Features
tomwalters@101: # Inner talkers
tomwalters@101: FEATURE_CLASS=aim
tomwalters@101: FEATURE_SUFFIX=$f
tomwalters@101: FEATURE_SIZE=12
tomwalters@101: FEATURE_TYPE=USER_E_D_A
tomwalters@101: TALKERS=outer_talkers
tomwalters@101: TRAINING_SNR=clean
tomwalters@101: run_train_test
tomwalters@101: ########################
tomwalters@101: done
tomwalters@91: 
tomwalters@78: 
tomwalters@76: 
tomwalters@81: