Maria@4: # -*- coding: utf-8 -*-
Maria@4: """
Maria@4: Created on Mon Oct 19 12:58:07 2015
Maria@4: 
Maria@4: @author: mariapanteli
Maria@4: """
Maria@4: import librosa
Maria@4: import scipy.signal
Maria@4: import numpy
Maria@4: 
Maria@4: class MFCCs:
Maria@4:     def __init__(self):
Maria@4:         self.y = None
Maria@4:         self.sr = None
Maria@4:         self.melspec = None
Maria@4:         self.melsr = None
Maria@4:         self.win1 = None
Maria@4:         self.hop1 = None
Maria@4:         self.mfccs = None
Maria@4: 
Maria@4:     def load_audiofile(self, filename='test.wav', sr=None, segment=True):
Maria@4:         self.y, self.sr = librosa.load(filename, sr=sr)
Maria@4:         if segment:
Maria@4:             tracklength = self.y.shape[0]/float(self.sr)
Maria@4:             startSample = 0
Maria@4:             endSample = None
Maria@4:             if tracklength > 90:
Maria@4:                 startPointSec = (tracklength/2.)-20
Maria@4:                 startSample = round(startPointSec*self.sr)
Maria@4:                 endSample = startSample+45*self.sr
Maria@4:             self.y = self.y[startSample:endSample]
Maria@4:     
Maria@4:     def mel_spectrogram(self, y=None, sr=None):
Maria@4:         if self.y is None:
Maria@4:             self.y = y
Maria@4:         if self.sr is None:
Maria@4:             self.sr = sr
Maria@4:         win1 = int(round(0.04*self.sr))
Maria@4:         hop1 = int(round(win1/8.))
Maria@4:         nfft1 = int(2**numpy.ceil(numpy.log2(win1)))
Maria@4:         nmels = 40
Maria@4:         D = numpy.abs(librosa.stft(self.y, n_fft=nfft1, hop_length=hop1, win_length=win1, window=scipy.signal.hamming))**2
Maria@4:         #melspec = librosa.feature.melspectrogram(S=D, sr=self.sr, n_mels=nmels)  
Maria@4:         melspec = librosa.feature.melspectrogram(S=D, sr=self.sr, n_mels=nmels, fmax=8000)  
Maria@4:         melsr = self.sr/float(hop1)
Maria@4:         self.melspec = melspec
Maria@4:         self.melsr = melsr
Maria@4:         
Maria@4:     def calc_mfccs(self, y=None, sr=None):
Maria@4:         if self.y is None:
Maria@4:             self.y = y
Maria@4:         if self.sr is None:
Maria@4:             self.sr = sr
Maria@4:         # require log-amplitude
Maria@4:         self.mfccs = librosa.feature.mfcc(S=librosa.logamplitude(self.melspec), n_mfcc=21)
Maria@4:         # remove DCT component        
Maria@4:         self.mfccs = self.mfccs[1:,:]
Maria@4:         
Maria@4:     def get_mfccs(self, filename='test.wav', secondframedecomp=False):
Maria@4:         self.load_audiofile(filename=filename)
Maria@4:         self.mel_spectrogram()
Maria@4:         self.calc_mfccs()
Maria@4:         
Maria@4:         if secondframedecomp:
Maria@4:             win2 = int(round(8*self.melsr))
Maria@4:             hop2 = int(round(0.5*self.melsr))
Maria@4:             nbins, norigframes = self.melspec.shape
Maria@4:             nframes = int(1+numpy.floor((norigframes-win2)/float(hop2)))
Maria@4:             avemfccs = numpy.empty((nbins, nframes))
Maria@4:             for i in range(nframes):  # loop over all 8-sec frames
Maria@4:                 avemfccs[:,i] = numpy.mean(self.mfccs[:, (i*hop2):(i*hop2+win2)], axis=1, keepdims=True)            
Maria@4:             self.mfccs = avemfccs
Maria@4:         return self.mfccs
Maria@4:     
Maria@4:     def get_mfccs_from_melspec(self, melspec=[], melsr=[]):
Maria@4:         self.melspec = melspec
Maria@4:         self.melsr = melsr
Maria@4:         self.calc_mfccs()
Maria@4:         return self.mfccs
Maria@4: 
Maria@4: if __name__ == '__main__':
Maria@4:     mfs = MFCCs()
Maria@4:     mfs.get_mfccs()