tomwalters@335: #!/bin/bash
tomwalters@335: # Copyright 2010 Thomas Walters <tom@acousticscale.org>
tomwalters@335: #
tomwalters@335: # Run a series of experiments which compare MFCC features generated by HTK to
tomwalters@335: # AIM features generated using AIM-C using a series of syllable recogntiton
tomwalters@335: # tasks.
tomwalters@371: # This script expects to be run from within the AIM-C source tree.
tomwalters@382: # It builds the HTK binaries and AIM-C AIMCopy binary if they're not
tomwalters@382: # present.
tomwalters@371: # The following environment varaibles should be set before this script is run:
tomwalters@371: # SYLLABLES_DATABASE_URL - URL of a tar file containing the CNBH syllables
tomwalters@371: # database in FLAC format
tomwalters@382: # HTK_USERNAME and HTK_PASSWORD - username and password for the site at
tomwalters@371: # http://htk.eng.cam.ac.uk/
tomwalters@379: # NUMBER_OF_CORES - total number of machine cores
tomwalters@371: 
tomwalters@413: sudo apt-get -y update
tomwalters@413: sudo apt-get -y install bc subversion scons pkg-config \
tomwalters@413:                 libsndfile1-dev build-essential libboost-dev \
tomwalters@414:                 python sox python-matplotlib libcairo-dev
tomwalters@413: 
tomwalters@413: # For 64-bit systems, uncomment this line:
tomwalters@413: sudo apt-get -y install libc6-dev-i386
tomwalters@413: 
tomwalters@335: # Set these to be the location of your input database, and desired output
tomwalters@382: # locations. (Note: the user running this script needs write permissions on
tomwalters@382: # the $WORKING_VOLUME.)
tomwalters@417: WORKING_VOLUME=/mnt/scratch0/aim
tomwalters@382: 
tomwalters@382: SYLLABLES_DATABASE_TAR=$WORKING_VOLUME/001-downloaded_sounds_data/cnbh-syllables.tar
tomwalters@382: SOUNDS_ROOT=$WORKING_VOLUME/002-sounds/
tomwalters@382: FEATURES_ROOT=$WORKING_VOLUME/003-features/
tomwalters@382: HMMS_ROOT=$WORKING_VOLUME/004-hmms/
tomwalters@382: HTK_ROOT=$WORKING_VOLUME/software/htk/
tomwalters@382: AIMC_ROOT=$WORKING_VOLUME/software/aimc/
tomwalters@335: 
tomwalters@416: THIS_DIR=`dirname $0`
tomwalters@415: AIMCOPY_CONFIGURATION_FILE=$THIS_DIR/cnbh-syllables/feature_generation/ssi_profile_features.aimcopycfg
tomwalters@413: 
tomwalters@335: # Number of cores on the experimental machine. Various scripts will try to use
tomwalters@335: # this if it's set.
tomwalters@379: # NUMBER_OF_CORES=8
tomwalters@335: 
tomwalters@335: # Fail if any command fails
tomwalters@335: set -e
tomwalters@335: 
tomwalters@335: # Fail if any variable is unset
tomwalters@335: set -u
tomwalters@335: 
tomwalters@382: ######
tomwalters@382: # Step 001 - Get the sounds database
tomwalters@348: if [ ! -e $SYLLABLES_DATABASE_TAR ]; then
tomwalters@382:   mkdir -p `dirname $SYLLABLES_DATABASE_TAR`
tomwalters@348:   wget -O $SYLLABLES_DATABASE_TAR $SYLLABLES_DATABASE_URL
tomwalters@348: fi
tomwalters@348: 
tomwalters@335: if [ ! -d $SOUNDS_ROOT ]; then
tomwalters@382:   mkdir -p $SOUNDS_ROOT
tomwalters@335: fi
tomwalters@335: 
tomwalters@382: # Untar the CNBH syllables database, and convert the files from FLAC to WAV.
tomwalters@335: if [ ! -e $SOUNDS_ROOT/.untar_db_success ]; then
tomwalters@335:   tar -x -C $SOUNDS_ROOT -f $SYLLABLES_DATABASE_TAR
tomwalters@335:   touch $SOUNDS_ROOT/.untar_db_success
tomwalters@335: fi
tomwalters@335: 
tomwalters@335: # Convert the database to .WAV format and place it in $SOUNDS_ROOT/clean
tomwalters@355: echo "Converting CNBH-syllables database from FLAC to WAV..."
tomwalters@341: ./cnbh-syllables/feature_generation/convert_flac_to_wav.sh $SOUNDS_ROOT
tomwalters@382: #
tomwalters@382: ######
tomwalters@335: 
tomwalters@382: #####
tomwalters@382: # Step 002 -
tomwalters@335: # Generate versions of the CNBH syllables spoke pattern with a range of
tomwalters@335: # signal-to-noise ratios (SNRs). The versions are put in the directory
tomwalters@335: # ${SOUNDS_ROOT}/${SNR}_dB/ for each SNR in $SNRS.
tomwalters@417: SNRS="45"#" 42 39 36 33" #" 30 27 24 21 18 15 12 9 6 3 0"
tomwalters@377: #SNRS="30" # For testing
tomwalters@356: ./cnbh-syllables/feature_generation/pink_noise.sh $SOUNDS_ROOT/clean/ "$SNRS"
tomwalters@335: 
tomwalters@335: # Make the list of all feature drectories
tomwalters@417: #FEATURE_DIRS="clean"
tomwalters@417: FEATURE_DIRS=""
tomwalters@335: for SNR in $SNRS; do
tomwalters@335:   FEATURE_DIRS="$FEATURE_DIRS snr_${SNR}dB"
tomwalters@335: done
tomwalters@335: 
tomwalters@335: # Generate feature sets (for the full range of SNRs in $FEATURE_DIRS)
tomwalters@335: # 1. Standard MFCC features
tomwalters@335: # 2. AIM features
tomwalters@382: # 3. MFCC features with optimal VTLN
tomwalters@339: 
tomwalters@339: if [ ! -d $FEATURES_ROOT ]; then
tomwalters@382:  mkdir -p $FEATURES_ROOT
tomwalters@339: fi
tomwalters@339: 
tomwalters@382: if [ ! -e $HTK_ROOT/.htk_installed_success ]; then
tomwalters@382:   ./HTK/install_htk.sh $HTK_ROOT
tomwalters@351: fi
tomwalters@350: 
tomwalters@382: if [ ! -e $AIMC_ROOT/.aimc_build_success ]; then
tomwalters@382:  ./aimc/build_aimc.sh $AIMC_ROOT
tomwalters@353: fi
tomwalters@415: export PATH=$PATH:$AIMC_ROOT/build/posix-release/
tomwalters@353: 
tomwalters@335: for SOURCE_SNR in $FEATURE_DIRS; do
tomwalters@347:   if [ ! -e $FEATURES_ROOT/mfcc/$SOURCE_SNR/.make_mfcc_features_success ]; then
tomwalters@335:     mkdir -p $FEATURES_ROOT/mfcc/$SOURCE_SNR/
tomwalters@335:     # Generate the list of files to convert
tomwalters@359:     ./cnbh-syllables/feature_generation/gen_hcopy_aimcopy_script.sh $FEATURES_ROOT/mfcc/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/ htk
tomwalters@335:     # Run the conversion
tomwalters@382:     ./cnbh-syllables/feature_generation/run_hcopy.sh $FEATURES_ROOT/mfcc/$SOURCE_SNR/ $NUMBER_OF_CORES
tomwalters@382:     touch $FEATURES_ROOT/mfcc/$SOURCE_SNR/.make_mfcc_features_success
tomwalters@346:   fi
tomwalters@335: 
tomwalters@347:   if [ ! -e $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/.make_mfcc_vtln_features_success ]; then
tomwalters@335:     mkdir -p $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/
tomwalters@335:     # Generate the file list and run the conversion (all one step, since this
tomwalters@356:     # version uses a different configuration for each talker)
tomwalters@382:     ./cnbh-syllables/feature_generation/run_mfcc_vtln_conversion.sh $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/
tomwalters@382:     touch $FEATURES_ROOT/mfcc_vtln/$SOURCE_SNR/.make_mfcc_vtln_features_success
tomwalters@346:   fi
tomwalters@335: 
tomwalters@347:   if [ ! -e $FEATURES_ROOT/aim/$SOURCE_SNR/.make_aim_features_success ]; then
tomwalters@382:     mkdir -p $FEATURES_ROOT/aim/$SOURCE_SNR/
tomwalters@359:     ./cnbh-syllables/feature_generation/gen_hcopy_aimcopy_script.sh $FEATURES_ROOT/aim/$SOURCE_SNR/ $SOUNDS_ROOT/$SOURCE_SNR/ ""
tomwalters@335:     # Run the conversion
tomwalters@413:     ./cnbh-syllables/feature_generation/run_aimcopy.sh $AIMCOPY_CONFIGURATION_FILE $FEATURES_ROOT/aim/$SOURCE_SNR/ $NUMBER_OF_CORES
tomwalters@379:     touch $FEATURES_ROOT/aim/$SOURCE_SNR/.make_aim_features_success
tomwalters@346:   fi
tomwalters@382: done
tomwalters@335: 
tomwalters@382: mkdir -p $HMMS_ROOT
tomwalters@363: 
tomwalters@335: # Now run a bunch of experiments.
tomwalters@335: # For each of the feature types, we want to run HMMs with a bunch of
tomwalters@335: # parameters.
tomwalters@382: TRAINING_ITERATIONS="0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15" # 16 17 18 19 20"
tomwalters@413: #TESTING_ITERATIONS="0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15" #" 16 17 18 19 20"
tomwalters@382: TESTING_ITERATIONS="15"
tomwalters@382: #HMM_STATES="3 4 5 6 7 8"
tomwalters@382: HMM_STATES="4"
tomwalters@382: #HMM_OUTPUT_COMPONENTS="1 2 3 4 5 6 7"
tomwalters@382: HMM_OUTPUT_COMPONENTS="4"
tomwalters@372: 
tomwalters@374: run_train_test () {
tomwalters@377: # TODO(tom): Make sure that the training SNR is generated first
tomwalters@374: for SOURCE_SNR in $FEATURE_DIRS; do
tomwalters@374: WORK=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/$SOURCE_SNR/$TALKERS/
tomwalters@374: mkdir -p $WORK
tomwalters@374: FEATURES_DIR=$FEATURES_ROOT/$FEATURE_CLASS/$SOURCE_SNR/
tomwalters@385: SPOKE_PATTERN_FILE=`pwd`/cnbh-syllables/run_training_and_testing/train_test_sets/gen_spoke_points/spoke_pattern.txt
tomwalters@374: 
tomwalters@374: ./cnbh-syllables/run_training_and_testing/train_test_sets/generate_train_test_lists.sh \
tomwalters@374:     $TALKERS \
tomwalters@374:     $WORK \
tomwalters@374:     $FEATURES_DIR \
tomwalters@374:     $FEATURE_SUFFIX
tomwalters@374: 
tom@442: if [ $TRAINING_SNR == 'random' ]; then
tom@442:   TRAINING_SCRIPT=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/clean/$TALKERS/training_script
tom@442:   TRAINING_MASTER_LABEL_FILE=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/clean/$TALKERS/training_master_label_file
tom@442:   RANDOMIZED_TRAINING_SCRIPT=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/clean/$TALKERS/training_script_randomized
tom@442:   RANDOMIZED_TRAINING_MASTER_LABEL_FILE=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/clean/$TALKERS/training_master_label_file_randomized
tom@442:   ./cnbh-syllables/run_training_and_testing/train_test_sets/randomize_snrs.py -s clean -f $TRAINING_SCRIPT -m $TRAINING_MASTER_LABEL_FILE -o $RANDOMIZED_TRAINING_SCRIPT -p $RANDOMIZED_TRAINING_MASTER_LABEL_FILE
tom@442:   TRAINING_SCRIPT=$RANDOMIZED_TRAINING_SCRIPT
tom@442:   TRAINING_MASTER_LABEL_FILE=$RANDOMIZED_TRAINING_MASTER_LABEL_FILE
tom@442: else
tom@442:   TRAINING_SCRIPT=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/$TRAINING_SNR/$TALKERS/training_script
tom@442:   TRAINING_MASTER_LABEL_FILE=$HMMS_ROOT/$FEATURE_CLASS/$FEATURE_SUFFIX/$TRAINING_SNR/$TALKERS/training_master_label_file
tom@442: fi
tomwalters@377: 
tomwalters@374: TESTING_SCRIPT=$WORK/testing_script
tomwalters@377: TESTING_MASTER_LABEL_FILE=$WORK/testing_master_label_file
tomwalters@374: 
tomwalters@374: ./cnbh-syllables/run_training_and_testing/gen_htk_base_files.sh $WORK
tomwalters@374: 
tomwalters@374: ./cnbh-syllables/run_training_and_testing/test_features.sh \
tomwalters@374:     "$WORK" \
tomwalters@374:     "$FEATURES_ROOT/$FEATURE_CLASS/$SOURCE_SNR/" \
tomwalters@374:     "$FEATURE_SUFFIX" \
tomwalters@374:     "$HMM_STATES" \
tomwalters@374:     "$HMM_OUTPUT_COMPONENTS" \
tomwalters@374:     "$TRAINING_ITERATIONS" \
tomwalters@374:     "$TESTING_ITERATIONS" \
tomwalters@374:     "$FEATURE_SIZE" \
tomwalters@374:     "$FEATURE_TYPE" \
tomwalters@374:     "$TRAINING_SCRIPT" \
tomwalters@374:     "$TESTING_SCRIPT" \
tomwalters@376:     "$TRAINING_MASTER_LABEL_FILE" \
tomwalters@385:     "$TESTING_MASTER_LABEL_FILE" \
tomwalters@385:     "$SPOKE_PATTERN_FILE"
tomwalters@374: done
tomwalters@374: }
tomwalters@374: 
tomwalters@372: ########################
tomwalters@372: # Standard MFCCs
tomwalters@356: FEATURE_CLASS=mfcc
tomwalters@360: FEATURE_SUFFIX=htk
tomwalters@356: FEATURE_SIZE=39
tomwalters@356: FEATURE_TYPE=MFCC_0_D_A
tomwalters@372: TALKERS=inner_talkers
tom@442: TRAINING_SNR=random
tomwalters@372: run_train_test
tomwalters@372: ########################
tomwalters@335: 
tomwalters@372: ########################
tomwalters@372: # Standard MFCCs
tomwalters@372: # Train on extrema
tomwalters@372: FEATURE_CLASS=mfcc
tomwalters@372: FEATURE_SUFFIX=htk
tomwalters@372: FEATURE_SIZE=39
tomwalters@372: FEATURE_TYPE=MFCC_0_D_A
tomwalters@372: TALKERS=outer_talkers
tomwalters@377: TRAINING_SNR=clean
tom@442: #run_train_test
tomwalters@372: ########################
tomwalters@372: 
tomwalters@372: ########################
tomwalters@372: # MFCCs with VTLN
tomwalters@372: FEATURE_CLASS=mfcc_vtln
tomwalters@372: FEATURE_SUFFIX=htk
tomwalters@372: FEATURE_SIZE=39
tomwalters@372: FEATURE_TYPE=MFCC_0_D_A
tomwalters@359: TALKERS=inner_talkers
tom@442: TRAINING_SNR=random
tomwalters@372: run_train_test
tomwalters@372: ########################
tomwalters@372: 
tomwalters@372: ########################
tomwalters@372: # MFCCs with VTLN
tomwalters@372: # Train on extrema
tomwalters@372: FEATURE_CLASS=mfcc_vtln
tomwalters@372: FEATURE_SUFFIX=htk
tomwalters@372: FEATURE_SIZE=39
tomwalters@372: FEATURE_TYPE=MFCC_0_D_A
tomwalters@372: TALKERS=outer_talkers
tom@442: TRAINING_SNR=random
tom@442: #run_train_test
tomwalters@372: ########################
tomwalters@372: 
tomwalters@382: AIM_FEATURE_SUFFIXES="slice_1_no_cutoff ssi_profile_no_cutoff slice_1_cutoff ssi_profile_cutoff smooth_nap_profile"
tomwalters@382: for f in $AIM_FEATURE_SUFFIXES
tomwalters@382: do
tomwalters@372: ########################
tomwalters@372: # AIM Features
tomwalters@382: # Inner talkers
tomwalters@382: FEATURE_CLASS=aim
tomwalters@382: FEATURE_SUFFIX=$f
tomwalters@382: FEATURE_SIZE=12
tomwalters@382: FEATURE_TYPE=USER_E_D_A
tomwalters@382: TALKERS=inner_talkers
tom@442: TRAINING_SNR=random
tomwalters@382: run_train_test
tomwalters@372: ########################
tomwalters@372: 
tomwalters@382: ########################
tomwalters@382: # AIM Features
tomwalters@382: # Inner talkers
tomwalters@382: FEATURE_CLASS=aim
tomwalters@382: FEATURE_SUFFIX=$f
tomwalters@382: FEATURE_SIZE=12
tomwalters@382: FEATURE_TYPE=USER_E_D_A
tomwalters@382: TALKERS=outer_talkers
tomwalters@382: TRAINING_SNR=clean
tom@442: #run_train_test
tomwalters@382: ########################
tomwalters@382: done
tomwalters@372: 
tomwalters@359: 
tomwalters@357: 
tomwalters@362: