# To control the system (exit the program, create new directories)
import sys
import os
# Keep track of the time --> time stamps; keep track of duration
import time
import datetime
# python library to handle matrices
import numpy as np
# import modules with functions I wrote to organize data and extract features
import feature_extraction_scripts.organize_speech_data as orgdata
import feature_extraction_scripts.feature_extraction_functions as featfun
# Handle exceptions I think might happen
from feature_extraction_scripts.errors import FeatureExtractionFail, ExitApp
##########################################################################
########################### FUNCTIONS ####################################
def get_date():
'''
This creates a string of the day, hour, minute and second
I use this to make folder names unique
For the files themselves, I generate genuinely unique names (i.e. name001.csv, name002.csv, etc.)
'''
time = datetime.datetime.now()
time_str = "{}y{}m{}d{}h{}m{}s".format(time.year,time.month,time.day,time.hour,time.minute,time.second)
return(time_str)
def main(data_path,feature_type,num_filters=None,delta=False,dom_freq=False,noise=False,vad=False,timesteps=None,context_window=None,noise_path=None,limit=None):
'''
Here defaults are set, if they aren't yet set by the user.
Number of Filters OR Number of Coefficients
FBANK or mel filterbank energies are produced based on the number of mel filters used.
*20 and 40 are pretty common.
MFCCs or mel frequency cepstral coefficients also use mel filters. With additional mathematic equations applied, to ultimately make speech easier for traditional machine learning algorithms to learn (they reduce the colinearity of the features)
* 13,20,40 coefficients are not uncommon.
'''
if num_filters is None and feature_type != "stft":
#Common number of coefficients / filters for MFCC and FBANK is 40
num_filters = 40
if feature_type == "stft":
#This number might change if you change the window size/ window shift
#settings within the feature extraction functions.
#I have put window size == 25ms and window shift == 10ms as default.
num_filters = 201
if noise is False:
#ensure that if the user doesn't want to add noise to data,
#that the wavefile variable is set to None
noise_path = None
if timesteps is None:
# For the LSTM: in how many segments do you want each recording to
# be split up? The LSTM would be fed each section, consecutively.
timesteps = 5
if context_window is None:
# This concerns how large you want each set of frames to be.
# The frame width contains a central frame, with surrounding context windows.
context_window = 5
#keep track of how many feature columns will be needed to store feature data
if delta:
#delta stands for both delta and delta delta
#these are the first and second derivatives of whatever features are extracted
#apparently these show 'rate of change' and 'rate of acceleration'
#and have been used in speech recognition/related machine learning tasks
num_feature_columns = num_filters * 3
else:
num_feature_columns = num_filters
if dom_freq is not False:
#dominant frequency is an experimental parameter... simply which
#frequency is the dominant one, at each of the sampled windows and window shifts.
num_feature_columns += 1
frame_width = context_window*2 + 1
#####################################################################
######################## HOUSE KEEPING ##############################
start = time.time()
time_stamp = get_date()
head_folder_beg = "./ml_speech_projects/"
curr_folder = "{}{}_models_{}".format(feature_type,num_feature_columns,time_stamp)
head_folder = head_folder_beg+curr_folder
#create folder to store all data (encoded labels, features)
if not os.path.exists(head_folder):
os.makedirs(head_folder)
'''
Collect all labels in data:
Labels should be the subdirectories of the data directory
Not included:
Folders/files with names:
* starting with "_"
* are typical GitHub files, like LICENSE
'''
try:
print("\nFeatures being collected: {}".format(feature_type.upper()))
print("\nTotal number of feature columns: {}".format(num_feature_columns))
labels_class = orgdata.collect_labels(data_path)
labels_print = ""
for i in range(len(labels_class)):
x = "\n{}) {}".format(i+1,labels_class[i])
labels_print += x
print("\nClasses found: {}\n".format(labels_print))
print("\nLimit: {}\n".format(limit))
print("\nIs this correct? (Y/N)")
correct_labels = input()
if 'y' in correct_labels.lower() or correct_labels.lower() == '':
pass
else:
raise ExitApp()
'''
Create labels-encoding dictionary:
This helps when saving data later to npy files
Integer encode the labels and save with feature data as label column
'''
dict_labels_encoded = orgdata.create_save_dict_labels_encode(labels_class,head_folder)
#create directories to store train, validation, and test data
train_val_test_filenames = []
train_val_test_directories = []
for i in ["train","val","test"]:
new_path = "{}/data_{}/".format(head_folder,i)
train_val_test_filenames.append(new_path+"{}_features".format(i))
train_val_test_directories.append(new_path)
try:
os.makedirs(new_path)
except OSError as e:
print("Directory ~ {} ~ already exists".format(new_path))
pass
#############################################
############## DATA ORGANIZATION ############
'''
I made the decision to balance the classes out. Real data doesn't have
balanced classes. But, because this code was prepared for a workshop,
one for exploring how feature extraction influenced deep learning models, I decided to remove that confounding factor.
'''
#collect filenames and labels of each filename
paths, labels_wavefile = orgdata.collect_audio_and_labels(data_path)
#to balance out the classes, find the label/class w fewest recordings
max_num_per_class, class_max_samps = orgdata.get_max_samples_per_class(labels_class,labels_wavefile)
'''
Create dictionary with labels and their indices in the lists: labels_wavefile and paths
useful in separating the indices into balanced train, validation, and test datasets
'''
dict_class_index_list = orgdata.make_dict_class_index(labels_class,labels_wavefile)
'''
Assign number of recordings for each dataset,
keeping the data balanced between classes
Defaults:
* .8 of max number of samples --> train
* .1 of max number of samples --> validation
* .1 of max number of samples --> test
'''
max_nums_train_val_test = orgdata.get_max_nums_train_val_test(max_num_per_class)
#randomly assign indices (evenly across class) to train, val, test datasets:
dict_class_dataset_index_list = orgdata.assign_indices_train_val_test(labels_class,dict_class_index_list,max_nums_train_val_test)
#make sure no indices mix between datasets:
orgdata.check_4_dataset_mixing(labels_class,dict_class_dataset_index_list)
except ExitApp:
sys.exit()
except IndexError:
print("\nError collecting data.\n".upper())
print("Double check how your data is organized. Are the labels printed above correct?\n")
sys.exit()
#############################################
############# FEATURE EXTRACTION ############
start_feature_extraction = time.time()
try:
for i in range(3):
dataset_index = i # 0 = train, 1 = validation, 2 = test
#see the script 'feature_extraction_functions.py' in the folder
#'feature_extraction_scripts' for more information about extraction.
extraction_completed = featfun.save_feats2npy(labels_class,dict_labels_encoded,train_val_test_filenames[dataset_index],max_nums_train_val_test[dataset_index],dict_class_dataset_index_list,paths,labels_wavefile,feature_type,num_filters,num_feature_columns,timesteps,frame_width,head_folder,limit=limit,delta=delta,dom_freq=dom_freq,noise_wavefile=noise_path,vad=vad,dataset_index=dataset_index)
if extraction_completed:
print("\nRound {} feature extraction successful.\n".format(i+1))
else:
print("\nRound {} feature extraction was unsuccessful.".format(i+1))
raise FeatureExtractionFail()
end_feature_extraction = time.time()
print("Duration of feature extraction: {} minutes".format(round((end_feature_extraction-start_feature_extraction)/60,2)))
#LOG THE SETTINGS OF FEATURE EXTRACTION IN CSV FILE
dict_info_feature_extraction = {"data path":data_path,"limit":limit,"features":feature_type,"num original features":num_filters,"num total features":num_feature_columns,"delta":delta,"dominant frequency":dom_freq,"noise":noise,"beginning silence removal":vad,"timesteps":timesteps,"context window":context_window,"num classes":len(labels_class),"time stamp":time_stamp, "duration in minutes":round((end_feature_extraction-start_feature_extraction)/60,2)}
orgdata.log_extraction_settings(dict_info_feature_extraction,head_folder)
orgdata.log_class4balance(max_num_per_class,class_max_samps,head_folder)
print("\nFolder name to copy and paste for the training script:".upper())
print("\n\n'{}'\n\n".format(curr_folder))
return True
except FeatureExtractionFail as e:
print(e)
print("Feature Extraction Error. Terminated feature extraction process.")
pass
return False
if __name__=="__main__":
#variables to set:
#which directory has the data?
data_path = "./data"
#should there be a limit on how many waves are processed?
limit = .05 # Options: False or fraction of data to be extracted
#which type of features to extract?
feature_type = "fbank" # "mfcc" "fbank" "stft"
#number of filters or coefficients? If STFT, doesn't matter.. can put None
num_filters = 40 # Options: 40, 20, 13, None
delta = False # Calculate the 1st and 2nd derivatives of features?
dom_freq = False # Kinda sorta... Pitch (dominant frequency)
noise = True # Add noise to speech data?
vad = True # Apply voice activity detection (removes the beginning and ending 'silence'/background noise of recordings)
timesteps = 5
context_window = 5
#If noise == True, put the pathway to that noise here:
noise_path = "./data/_background_noise_/doing_the_dishes.wav"
main(
data_path,feature_type,
num_filters=num_filters,delta=delta,noise=noise,vad=vad,dom_freq=dom_freq,
timesteps=timesteps,context_window=context_window,noise_path=noise_path,limit = limit
)
