get_mfcc.py

'''

takes a wav file as input and outputs its 13 co-efficient MFCC.
Only implemented for single path, single MFCC extraction and does not automatically work on whole directory.
'''

import numpy as np
import os
from numpy.linalg import norm
import librosa.display
from get_spectraldata import *
import matplotlib.pyplot as plt
from normalize_data import normalizeSoundData
from import_words import shuffle_in_unison_scary
from dtw import dtw

pre_emphasis = 0.97
frame_size = 0.025
frame_stride = 0.01
NFFT = 512
nfilt = 40
num_ceps = 12
cep_lifter = 22 # 0 for no filter
nonllpath = '../LL_chunks/chunk1.wav'
llpath = '../nonLL_chunks/chunk1.wav'


def librosaMfcc(path):
    list = absoluteFilePaths(path)
    data = []


    for file in list:
        mfcc = average(findMfcc(file))
        # deltah = delta(mfcc)
        #temp = np.concatenate((mfcc, deltah))
        data.append(mfcc)



    return np.array(data)

def plotMfcc(mfcc1, mfcc2):

    plt.subplot(1, 2, 1)
    librosa.display.specshow(mfcc1)
    plt.subplot(1, 2, 2)
    librosa.display.specshow(mfcc2)
    plt.show()



def computeDistace(mfcc1, mfcc2):

    print("Finding DTW between the 2 mfccs")
    dist, cost, acc_cost, path = dtw(mfcc1.T, mfcc2.T, dist=lambda x, y: norm(x - y, ord=1))
    print ('Normalized distance between the two sounds:', dist)
    plt.imshow(cost.T, origin='lower', cmap=plt.get_cmap('gray'), interpolation='nearest')
    plt.plot(path[0], path[1], 'w')
    plt.xlim((-0.5, cost.shape[0] - 0.5))
    plt.ylim((-0.5, cost.shape[1] - 0.5))
    plt.xlabel('MFCC Column Index (FET)')
    plt.ylabel('MFCC Column Index (Non-FET)')
    plt.show()

def padMfcc(mfcc, fixedsize = 30):
    print ("padding mfcc")
    np.transpose(mfcc)
    # mfcc too small
    diff = fixedsize - mfcc.shape[1]
    print (diff)
    for _ in range(diff):
        print ("running")
        np.vstack((mfcc, np.zeros(20)))

    print (mfcc.shape)

def getMfccAverage(): # return average of all mfccs for each word in numpy array

    # first creating the dir list
    pathlist = []

    pathlist.extend(absoluteFilePaths('../LL_chunks'))
    pathlist.extend(absoluteFilePaths('../nonLL_chunks'))

    print ("path list is",pathlist)

    data = []

    for path in pathlist:
        print ("finding mfcc of", path)

        data.append(average(findMfcc(path)))

        # print len(data[-1])

    return np.array(data)

def getndimMfcc(): # return average of all mfccs for each word

    # first creating the dir list
    pathlist = []

    pathlist.extend(absoluteFilePaths('../LL_chunks'))
    pathlist.extend(absoluteFilePaths('../nonLL_chunks'))

    print ("path list is",pathlist)

    data = []

    for path in pathlist:
        print ("finding mfcc of", path)

        data.append(findMfcc(path))
        print(len(data[-1]))

    return np.array(data)

def getMfccDelta():
    # first creating the dir list
    print ("hello")
    pathlist = []

    pathlist.extend(absoluteFilePaths('../LL_chunks'))
    pathlist.extend(absoluteFilePaths('../nonLL_chunks'))

    print ("path list is", pathlist)

    data = []

    for path in pathlist:
        print ("finding mfcc of", path)
        # mfcc = average((findMfcc(path)))
        deltah = average(delta(findMfcc(path)))
        # temp = np.concatenate((deltah, mfcc))
        print (deltah)
        data.append(deltah)

        print (len(data[-1]))

    return np.array(data)



def absoluteFilePaths(directory):
   for dirpath,_,filenames in os.walk(directory):
       if ('.DS_Store' in filenames):
        filenames.remove('.DS_Store')
       for f in filenames:
           yield os.path.abspath(os.path.join(dirpath, f))



def findMfcc(path):
    y1, sr1 = librosa.load(path)
    mfcc = librosa.feature.mfcc(y1, sr1)
    return mfcc


def average(mfcc):
    # take an n coefficient mfcc for multiple samples and finds its nx1 size average array
    # print ("averaging mfcc")
    ave = []
    for i in range(mfcc.shape[0]):
        ave.append(np.mean(mfcc[i,:]))
    ave_numpy = np.array(ave)
    # print (ave_numpy)

    return ave_numpy

def getFinalNormalizedMfcc(): #shuffling occurs here

    data = getMfccAverage()
    # data = deltaplusmfcc()
    # data = getMfccSum()
    # data = getMfccDelta()

    normalizeSoundData(data)


    _, labels = getTrainingData()

    print (labels)

    shuffle_in_unison_scary(data, labels)


    return data, labels


def flatten():
    print("flattening")


def delta(mfcc):
    '''

    :param mfcc: input mfcc 2d array
    :return: first dervative of mfcc 2d array
    '''
    print ("finding the delta")
    return librosa.feature.delta(mfcc)

def sum(mfcc): #over time axis or coloum wise

    data = np.array(mfcc) # if mfcc data is not already in numpy

    return data.sum(axis=1)

def getlabels():
    '''

    :return: gets non shuffled labels as numpy array of size [training samples, 1], 1 for LL 0 for Non LL
    '''


def getMfccSum():
    pathlist = []

    pathlist.extend(absoluteFilePaths('../LL_chunks'))
    pathlist.extend(absoluteFilePaths('../nonLL_chunks'))

    print ("path list is", pathlist)

    data = []

    for path in pathlist:
        print ("finding mfcc of", path)

        data.append(sum(findMfcc(path)))

        print (len(data[-1]))

    return np.array(data)




def getmfccDoubleDelta():
    print






if __name__ == '__main__':

    # nonllmfcc, llmfcc = librosaMFCC(nonllpath, llpath)
    # average(llmfcc)

    # a = []
    #
    # for i in range(5):
    #     temp = findMfcc('../nonLL_chunks/chunk{}.wav'.format(i+1))
    #     np.savetxt('nonll{}.csv'.format(i+1), temp)

    mfcc1 = findMfcc('../tempsentences/fox.wav')
    mfcc2 = findMfcc('../tempsentences/foxn.wav')
    print(mfcc1.shape)
    print(mfcc2.shape)
    computeDistace(mfcc1, mfcc2)



    # plotMfcc(nonllmfcc, llmfcc)
    # computeDistace(nonllmfcc, llmfcc)


    # data, labels = getFinalNormalizedMfcc()
    # print (data[1:10, :])
    # print (data.size)
    # print (type(data))
    # # print data[500:,:]
    #
    # # deltaplusmfcc()