def sparkFeatureExt(line):
print line
string1 = file
cut = string1[72:]
cut = cut[:-10]
intClass = ClassToInt(cut)
(samplerate, wavedata) = wavfile.read(file)
(s1, n1) = spectral_centroid(wavedata, 512, samplerate)
(sr1, nr1) = spectral_rolloff(wavedata, 512, samplerate)
(sf1, nf1) = spectral_flux(wavedata, 512, samplerate)
(rms, ts) = root_mean_square(wavedata, 512, samplerate)
rms = rms[~np.isnan(
rms
)] #rms array contains NAN values and we have to remove these values
(zcr, ts1) = zero_crossing_rate(wavedata, 512, samplerate)
(MFCCs, mspec, spec) = mfcc(wavedata)
MFCC_coef = list()
ran = MFCCs.shape
ran1 = ran[0]
for ind1 in range(13):
sum = 0
summ = 0
for ind in range(ran1):
sum += MFCCs[ind, ind1]
MFCC_coef.append(sum / ran1)
eng = stEnergy(wavedata)
#Win = 0.050
#Step = 0.050
#eps = 0.00000001
return s1, sr1, sf1, rms, zcr, eng, MFCC_coef, intClass
def apply(self, data, meta=None):
all_ceps = []
for ch in data:
ceps, mspec, spec = mfcc(ch)
all_ceps.append(ceps.ravel())
return to_np_array(all_ceps)
def extract_mfcc(full_audio_path):
sample_rate, wave = wavfile.read(full_audio_path)
mfcc_features = mfcc(wave,
nwin=int(sample_rate * 0.03),
fs=sample_rate,
nceps=12)[0]
return mfcc_features
def apply(self, data, meta=None):
all_ceps = []
for ch in data:
ceps, mspec, spec = mfcc(ch)
all_ceps.append(ceps.ravel())
return to_np_array(all_ceps)
def single():
# import wav files
files = import_wav_data_in_dir(datapath)
# prepare for training
# get feature vectors by mfcc
X = []
y = []
for f in files:
x, sample_rate = sf.read(datapath + f)
x = np.clip(x, 1e-10, 1)
ceps, mspec, spec = mfcc(x, nwin=256, nfft=512, fs=8000, nceps=13)
X.append(np.mean(ceps, axis=0))
if '-hu.' in f:
y.append(label_dict['hu'])
elif '-ti.' in f:
y.append(label_dict['ti'])
else:
y.append(label_dict['dc'])
X = np.array(X)
y = np.array(y)
# training
clf = RandomForestClassifier(n_estimators=498, random_state=random_state)
# clf = XGBClassifier(max_depth=8, learning_rate=0.05, n_estimators=700, seed=random_state)
clf.fit(X, y)
# save model
joblib.dump(clf, 'trained_models/clf_rf2.pkl.cmp', compress=True)
def BED_extract(path, nfft):
list_data = numpy.array([])
list_label = numpy.array([])
"""
dic = {'W':[1,0],'L':[0,1],'E':[0,1],'A':[0,1],'F':[1,0],'T':[0,1],'N':[0.5,0.5]}
"""
dic = {'W':[0,1],'L':[0,1],'E':[0,1],'A':[0,1],'F':[1,0],'T':[0,1],'N':[0.5,0.5]}
for root, dir, files in os.walk(path):
rootpath = os.path.join(os.path.abspath(path), root)
for file in files:
if os.path.splitext(file)[1].lower()=='.wav':
filepath = os.path.join(rootpath, file)
SR, X = wavfile.read(filepath)
_, _, spec = mfcc(X, fs=SR, nfft=(nfft*2))
list_data = numpy.append(list_data, numpy.mean(spec, axis=0)[:nfft]/numpy.max(spec))
list_label = numpy.append(list_label, dic[file[5]])
list_data = numpy.reshape(list_data, (len(list_data)/nfft, nfft))
list_label = numpy.reshape(list_label, (len(list_label)/label_length, label_length))
return list_data, list_label
def apply(self, data):
all_ceps = []
for ch in data:
ceps, mspec, spec = mfcc(ch)
all_ceps.append(ceps.ravel())
return np.array(all_ceps)
def compute_features(source, features):
"""
compute features for all the tracks
"""
for label in source.keys():
for i in range(0,100):
base_path = os.path.join(FT_DIR, "%s_%d" % (label, i))
ft = []
if 'zcr' in features:
zcr, ts = zero_crossing_rate(source[label][i]['wavedata'], 512, source[label][i]['sample_rate'])
ft.append(zcr)
if 'rms' in features:
rms, ts = root_mean_square(source[label][i]['wavedata'], 512, source[label][i]['sample_rate'])
ft.append(rms)
if 'sc' in features:
sc, ts = spectual_centroid(source[label][i]['wavedata'], 2048, source[label][i]['sample_rate'])
ft.append(sc)
if 'sr' in features:
sr, ts = spectral_rolloff(source[label][i]['wavedata'], 2048, source[label][i]['sample_rate'])
ft.append(sr)
if 'sf' in features:
sf, ts = spectral_flux(source[label][i]['wavedata'], 2048, source[label][i]['sample_rate'])
ft.append(sf)
if 'mfcc' in features:
ceps, mspec, spec = mfcc(source[label][i]['wavedata'])
ft.append(ceps)
write_features(ft, base_path)
def calc_mfcc(data, fs):
mfcc_data, trush, trush = mfcc(data, nwin=256, nfft=512, fs=fs, nceps=13)
meanceps = np.zeros(mfcc_data[0].size)
for mc in mfcc_data:
meanceps += mc
return meanceps / mfcc_data[0].size
def extractFeature(self):
''' mfcc feature extraction '''
self.features = mfcc(self.sound,
nwin=int(self.soundSamplerate * 0.03),
fs=self.soundSamplerate,
nceps=13)[0]
def create_test_ceps():
"""
Creates the MFCC features from the test files,
saves them to disk, and returns the saved file name.
"""
for subdir, dirs, files in os.walk(TEST_DATASET_DIR):
genre = subdir[subdir.rfind('/', 0) + 1:]
print(genre)
if genre in genre_list:
count = 0
genre_ceps = np.zeros((30, 13), dtype=float)
print(subdir)
for file in files:
path = subdir + '/' + file
#print path
if path.endswith("wav"):
#print path
#create_ceps(path)
sample_rate, X = scipy.io.wavfile.read(path)
ceps, mspec, spec = mfcc(X)
num_ceps = len(ceps)
ceps = np.mean(ceps[int(num_ceps / 10):int(num_ceps * 9 /
10)],
axis=0)
genre_ceps[count] = ceps
count = count + 1
print(count)
#genre_ceps = np.array(genre_ceps)
print(genre_ceps.shape)
#break
write_ceps(genre_ceps, path)
def set_mfcc_matrix(self):
self.mfcc_matrix = mfcc(self.signal,
nwin=int(self.sample_rate * 0.03),
fs=self.sample_rate,
nceps=13)[0]
self.mfcc_matrix = self.mfcc_matrix[~np.isnan(self.mfcc_matrix).any(
axis=1)]
def load_validation_set():
"""
Output
a tuple of features: (fft features, mfcc features, mean-std features)
Description
extracts three types of features from validation set.
"""
ffts = dict()
mfccs = dict()
mean_stds = dict()
for i in validation_ids:
path = './validation/validation.{i}.wav'.format(i=i)
_, X = read_wav(path)
# FFT
fft = np.array(abs(sp.fft(X)[:1000]))
ffts.update({i: fft})
# MFCC
ceps, mspec, spec = mfcc(X)
num_ceps = len(ceps)
x = np.mean(ceps[int(num_ceps*1/10):int(num_ceps*9/10)], axis=0)
mfccs.update({i: x})
# Mean-Std
[Fs, x] = audioBasicIO.readAudioFile(path);
F = audioFeatureExtraction.stFeatureExtraction(x, Fs, 0.050*Fs, 0.025*Fs);
mean_std = []
for f in F:
mean_std.extend([f.mean(), f.std()])
mean_stds.update({i: np.array(mean_std)})
return (ffts, mfccs, mean_stds)
def mfccnceps(filenam):
print("creating ceps")
sample_rate, X = scipy.io.wavfile.read(filenam)
ceps, mspec, spec = mfcc(X)
basename, extn = os.path.splitext(filenam)
datafile = basename + ".ceps"
np.save(datafile, ceps) # cache results so that ML becomes fast
def classify(file_name):
accents = [
"CH",
"EN",
"IN",
"IR",
"IT",
"JA",
"KO",
]
models = glob("models/model*.xml")
models.sort()
if len(models) < 1:
print "no models found"
exit()
print "using model: {}".format(models[-1])
net = NetworkReader.readFrom(models[-1])
sample_rate, X = scipy.io.wavfile.read(file_name)
ceps, mspec, spec = mfcc(X)
x = []
num_ceps = len(ceps)
x.append(np.mean(ceps[int(num_ceps / 10):int(num_ceps * 9 / 10)], axis=0))
vx = np.array(x)
result = net.activate(vx[0].tolist()).tolist()
# print result
accent = accents[result.index(max(result))]
return accent
def BED_extract(path, nfft=784):
list_data = numpy.array([])
list_label = numpy.array([])
#W:anger:0 N:neutral:1 F:happiness:2 T:sadness:3
dic = {'W':0,'L':1,'E':3,'A':0,'F':2,'T':3,'N':1}
for root, dir, files in os.walk(path):
rootpath = os.path.join(os.path.abspath(path), root)
for file in files:
if os.path.splitext(file)[1].lower()=='.wav':
filepath = os.path.join(rootpath, file)
SR, X = wavfile.read(filepath)
_, _, spec = mfcc(X, fs=SR, nfft=(nfft*2))
print(filepath)
list_data = numpy.append(list_data, numpy.mean(spec, axis=0)[:nfft]/numpy.max(spec))
list_label = numpy.append(list_label, int(dic[file[5]]))
list_data = numpy.reshape(list_data, (len(list_data)/nfft, nfft))
return list_data, list_label
def apply(self, data):
all_ceps = []
for ch in data:
ceps, mspec, spec = mfcc(ch)
all_ceps.append(ceps.ravel())
return np.array(all_ceps)
def mfccIFY(dict_read):
dict = {}
for each in dict_read.keys():
[sample_rate,X] = dict_read.get(each)
ceps, mspec, spec = mfcc(X)
dict[each] = ceps
return dict
def emsenble():
# import wav files
files = import_wav_data_in_dir(datapath)
# prepare for training
# get feature vectors by mfcc
X = []
y = []
for f in files:
x, sample_rate = sf.read(datapath + f)
x = np.clip(x, 1e-10, 1)
ceps, mspec, spec = mfcc(x, nwin=256, nfft=512, fs=8000, nceps=13)
X.append(np.mean(ceps, axis=0))
if '-hu.' in f:
y.append(label_dict['hu'])
elif '-ti.' in f:
y.append(label_dict['ti'])
else:
y.append(label_dict['dc'])
X = np.array(X)
y = np.array(y)
# training
clf1 = RandomForestClassifier(n_estimators=498, random_state=random_state)
clf2 = KNeighborsClassifier(n_neighbors=3, weights='uniform', p=1)
clf3 = QuadraticDiscriminantAnalysis()
eclf = VotingClassifier(estimators=[('rf', clf1), ('knn', clf2),
('qda', clf3)],
voting='hard')
eclf.fit(X, y)
# save model
joblib.dump(eclf, 'trained_models/clf_rf_knn_qda.pkl.cmp', compress=True)
def getmfccdata(path):
"""
This function extracts the mfcc data from the wav files
Parameters:
-----------
path - path to get the directory name of the songs present "E:\UNM\CS 529 - Intro to Machine Learning\Assignment 3\opihi.cs.uvic.ca\sound\genres"
Returns:
--------
mfccdata - mfcc data matrix of size (600,13)
"""
classesmatrix = np.zeros((no_of_docs, 1)) # Stores the song, genre information in classesmatrix.txt file -> Line number as song index, genre
mfccdata = np.zeros((no_of_docs, no_of_mfcc_features)) # Matrix (600,13) to store the fft features information of all the songs in 6 genres
fileindex = 0 # to store the current offset of the song
for subdir, dirs, files in os.walk(path): # Traversing all the files in 6 genres
if os.path.basename(subdir) in genres.keys():
for f in files:
if f.endswith('.wav'):
print "Processing file : " + f
sample_rate, X = scipy.io.wavfile.read(os.path.join(subdir, f))
ceps, mspec, spec = mfcc(X)
num_ceps = ceps.shape[0]
mfcc_features = np.mean(ceps[int(num_ceps * 1 / 10):int(num_ceps * 9 / 10)], axis=0) # Extracts 13 features.
for i in range(len(mfcc_features)):
mfccdata[fileindex][i] = mfcc_features[i]
classesmatrix[fileindex] = genres[os.path.basename(subdir)] # Storing the genre of every song in a matrix.
fileindex += 1
np.savetxt('classesmatrix.txt', classesmatrix, '%d') # Writing the classesmatrix to a file.
return mfccdata
def create_ceps(fn):
#print(">>>: [%s]" % fn)
sample_rate, X = scipy.io.wavfile.read(fn)
#print("> - [%d]" % sample_rate)
ceps, mspec, spec = mfcc(X,nceps=13)
write_ceps(ceps, fn)
def sparkFeatureExt(line):
print line
string1= file
cut=string1[72:]
cut=cut[:-10]
intClass=ClassToInt(cut)
(samplerate, wavedata) = wavfile.read(file)
(s1,n1)= spectral_centroid(wavedata,512,samplerate)
(sr1,nr1)= spectral_rolloff(wavedata,512,samplerate)
(sf1,nf1)= spectral_flux(wavedata,512,samplerate)
(rms,ts) = root_mean_square(wavedata, 512, samplerate);
rms= rms[~np.isnan(rms)] #rms array contains NAN values and we have to remove these values
(zcr,ts1) = zero_crossing_rate(wavedata, 512, samplerate);
(MFCCs, mspec, spec) = mfcc(wavedata)
MFCC_coef=list()
ran=MFCCs.shape
ran1=ran[0]
for ind1 in range(13):
sum=0
summ=0
for ind in range(ran1):
sum+=MFCCs[ind,ind1]
MFCC_coef.append(sum/ran1)
eng= stEnergy(wavedata)
#Win = 0.050
#Step = 0.050
#eps = 0.00000001
return s1,sr1,sf1,rms,zcr,eng,MFCC_coef,intClass
def create_ceps(fn):
"""
Creating the MFCC features.
"""
sample_rate, X = scipy.io.wavfile.read(fn)
X[X == 0] = 1
ceps, mspec, spec = mfcc(X)
write_ceps(ceps, fn)
def mfccnceps(filenam):
print("creating ceps")
sample_rate, X = scipy.io.wavfile.read(filenam)
ceps, mspec, spec = mfcc(X)
basename, extn = os.path.splitext(filenam)
datafile = basename + ".ceps"
np.save(datafile, ceps)
return basename
def create_ceps(fn):
"""
Creates the MFCC features.
"""
sample_rate, X = scipy.io.wavfile.read(fn)
X[X==0]=1
ceps, mspec, spec = mfcc(X)
write_ceps(ceps, fn)
def generateMfcc(wavFile):
filteredFile = filtering(wavFile, 2800, 3400)
audio, fs, enc = wavread(filteredFile)
size = getFrameSize(filteredFile)
ceps, mspec, spec = mfcc(audio, nwin=size, nfft=size, fs=fs, nceps=13)
return ceps
def mfcc_sound_features(self):
ceps, mspec, spec = mfcc(self.audio)
num_ceps = len(ceps)
#v = np.mean(ceps[int(num_ceps / 10):int(num_ceps * 9 / 10)], axis=0)
v = np.mean(ceps, axis=0)
return v
def coeff(arr): ceps, mspec, spec = mfcc(arr,fs=11025) #print np.shape(ceps) auxceps = np.zeros(len(ceps)*13) for i in range(0,len(ceps)): for j in range(0,13): auxceps[13*i+j]=ceps[i][j] return(auxceps)
def convert(path):
data = {}
data["sample_rate"], X = scipy.io.wavfile.read(path)
data["ceps"], data["mspec"], data["spec"] = mfcc(X) #save everything it gives us in case it's useful lmao
cep_count = len(data["ceps"])
input_vector = np.array([np.mean(data["ceps"][int(cep_count / 10):int(cep_count * 9 / 10)], axis=0)])
return input_vector
def create_ceps(fn):
print "creating :", fn
sample_rate, X = w.read(fn)
ceps, mspec, spec = mfcc(X)
num_ceps = len(ceps)
return [
np.mean(ceps[int(num_ceps * 1 / 10):int(num_ceps * 9 / 10)], axis=0)
]
def mfcc_sound_features(self):
ceps, mspec, spec = mfcc(self.audio)
num_ceps = len(ceps)
#v = np.mean(ceps[int(num_ceps / 10):int(num_ceps * 9 / 10)], axis=0)
v = np.mean(ceps, axis=0)
return v
def show(self, example):
sound = audiolab.sndfile(self.base + example.file)
frames = sound.read_frames(sound.get_nframes()) * 0.8
mfcc = features.mfcc(frames[example.start:example.stop:2], fs=41000)
print mfcc[0].shape
fig = plt.figure()
fig.set_size_inches(20, 20)
ax = fig.add_subplot(111)
ax.imshow(mfcc[0].transpose()[:, :100])
def MFCC(package):
if type(package) is eT.EEGpackage:
all_ceps = []
for ch in package.packet:
ceps, mspec, spec = mfcc(ch) #mfcc not defined
all_ceps.append(ceps.ravel())
return np.array(all_ceps)
else:
return False
def create_ceps(wavfile): sampling_rate, song_array = scipy.io.wavfile.read(wavfile) """Get MFCC ceps : ndarray of MFCC mspec : ndarray of log-spectrum in the mel-domain spec : spectrum magnitude """ ceps, mspec, spec = mfcc(song_array) write_ceps(ceps, wavfile)
def create_ceps(fn):
sample_rate, X = io.wavfile.read(fn)
ceps, mspec, spec = mfcc(X)
isNan = False
for num in ceps:
if np.isnan(num[1]):
isNan = True
if isNan == False:
write_ceps(ceps, fn)
def performMFCC(self): sample_rate, X = scipy.io.wavfile.read(self.filename) X[X==0]=1 ceps, mspec, spec = mfcc(X) self.write_ceps(ceps) plt.plot(ceps) #plt.show() plt.title(self.filename) plt.savefig(os.path.splitext(self.filename)[0]+".png")
def show(self, example): sound = audiolab.sndfile(self.base + example.file) frames = sound.read_frames(sound.get_nframes()) * 0.8 mfcc = features.mfcc(frames[example.start: example.stop:2], fs=41000) print mfcc[0].shape fig = plt.figure() fig.set_size_inches(20, 20) ax = fig.add_subplot(111) ax.imshow(mfcc[0].transpose()[:, :100])
def get_ceps(filepath):
"""
:param filepath: takes wav audio file path
:return: returns first 13 Mel-frequency cepstral coefficients
"""
sample_rate, X = scipy.io.wavfile.read(filepath)
ceps, mspec, spec = mfcc(X)
ceps_ = np.mean(ceps, axis=0)
return ceps_
def create_ceps(wavfile):
sampling_rate, song_array = scipy.io.wavfile.read(wavfile)
"""Get MFCC
ceps : ndarray of MFCC
mspec : ndarray of log-spectrum in the mel-domain
spec : spectrum magnitude
"""
ceps, mspec, spec = mfcc(song_array)
write_ceps(ceps, wavfile)
def generate_and_save_ceps(wave_file):
sample_rate, X = sp.io.wavfile.read(wave_file)
ceps, mspec, spec = mfcc(X) #default 13 coefficients
#ceps is a 2d array #frames*#coefficients
num_frames = len(ceps)
ceps_average_over_frames = np.mean(
ceps[int(num_frames * 0.1):int(num_frames * 0.9)],
axis=0) # leave starting one-tenth's and last one-tenth's
ceps_file = wave_file[:-3] + "ceps"
np.save(ceps_file, ceps_average_over_frames)
def create_ceps(fn, op='plain'): sample_rate, X = wavfile.read(fn) if op == 'norm': norm = StandardScaler() X = norm.fit_transform(X) ceps, mspec, spec = mfcc(X) write_ceps(ceps, fn, op)
def get_features(filename,feature_type):
if feature_type == 'mfcc':
test_audio_file = wave.open(filename, 'r')
fileContents = get_audio(test_audio_file)
# get mfcc coeffs and don't look at first (corresponds to energy in signal)
coeffs = mfcc(fileContents,fs=44100)[0][1:] # [1] is mel coeffs, and [2] is entire FFT data???
test_audio_file.close()
elif feature_type == 'nlse':
coeffs = timbrespace(filename) # default hopsize=3, i.e. 50%, i.e. 150 ms
return coeffs
def extractfeatures():
genredirs = sorted(os.listdir(GTZAN_PATH))
for dirname in genredirs:
files = sorted(glob.glob(GTZAN_PATH + '/' + dirname + '/' + '*.wav'))
for filename in files:
f = filename
[_, data] = scipy.io.wavfile.read(f)
ceps, _, _ = mfcc(data)
fceps = FEAT_DATA_PATH + '/' + dirname + '/' + os.path.basename(
filename) + '.ceps'
numpy.save(fceps, ceps)
def get_mfcc2(filename):
from scikits.talkbox.features import mfcc
import os
import scipy
rate, signal = scipy.io.wavfile.read(filename)
ceps, mspec, spec = mfcc(signal)
base_filename, ext = os.path.splitext(filename)
data_filename = base_filename + ".ceps"
np.save(data_filename, ceps)
print(" Written %s" % data_filename)
def update(self, step):
self.end_of_sample = False
self.frames = self.state.read_frames(step)
if self.frames is None:
if isinstance(self.state, PlaySample):
self.end_of_sample = True
self.state = self.state.next()
self.frames = self.state.read_frames(step)
assert self.frames is not None, "There are must be some frames"
self.mfcc = mfcc(self.frames,
fs=len(self.frames)/step, nceps=13)[0][0]
def get_mfcc2(filename):
from scikits.talkbox.features import mfcc
import os
import scipy
rate, signal = scipy.io.wavfile.read(filename)
ceps, mspec, spec = mfcc(signal)
base_filename, ext = os.path.splitext(filename)
data_filename = base_filename + ".ceps"
np.save(data_filename, ceps)
print(" Written %s" % data_filename)
def create_ceps():
for genre in GENRES:
wav_list = glob.glob(os.path.join(BASE_DIR, genre, '*.' + 'wav'))
npy_list = glob.glob(os.path.join(BASE_DIR, genre, '*.' + 'npy'))
npy_list = [name[:-9] for name in npy_list]
for fn in wav_list:
if fn[:-4] not in npy_list:
sample_rate, X = scipy.io.wavfile.read(fn)
ceps, mspec, spec = mfcc(X)
write_ceps(ceps, fn)
print 'created ', fn[:-4] + '.ceps.npy'
def generate_mfcc(path,test = False):
"""
generating each song's fft
"""
sample_rate, X = wavfile.read(path)
ceps, mspec, spec = mfcc(X)
if test==True:
print "writing test data"
write_mfcc(path,ceps,True)
else:
write_mfcc(path,ceps)
def create_fft(fn, myclass):
try:
sample_rate, X = scipy.io.wavfile.read(fn)
except ValueError:
return
ceps, mspec, spec = mfcc(X)
num_ceps = len(ceps)
x = np.mean(ceps[int(num_ceps * 1 / 10):int(num_ceps * 9 / 10)], axis=0)
y = np.int(myclass)
XA.append(x)
ya.append(y)
def show_specgram(speech):
sound = audiolab.sndfile(speech, 'read')
sound_info = sound.read_frames(sound.get_nframes())
#spectrogram = plt.specgram(sound_info)
mfcc = talkfeat.mfcc(sound_info)
#print mfcc
plt.imshow(mfcc[0].transpose())
plt.title('Spectrogram of %s' % sys.argv[1])
plt.show()
sound.close()
def readwav(trainfolder,testfolder):
os.chdir(trainfolder)
for f in glob.glob("*.wav"):
speaker,letter,_ = f.split('.')[0].split('-')
mfccname = f.split('.')[0]+".mfc"
data, fs = wavread(f)[:2]
cep= mfcc(data, fs=fs, nwin=int(fs*0.025))[0]
np.savetxt(mfccname,cep,fmt='%.10f')
os.chdir(testfolder)
for f in glob.glob("*.wav"):
speaker,letter,_ = f.split('.')[0].split('-')
mfccname = f.split('.')[0]+".mfc"
data, fs = wavread(f)[:2]
cep= mfcc(data, fs=fs, nwin=int(fs*0.025))[0]
np.savetxt(mfccname,cep,fmt='%.10f')
def create_ceps_test(fn):
"""
Creates the MFCC features from the test files,
saves them to disk, and returns the saved file name.
"""
sample_rate, X = scipy.io.wavfile.read(fn)
X[X==0]=1
np.nan_to_num(X)
ceps, mspec, spec = mfcc(X)
base_fn, ext = os.path.splitext(fn)
data_fn = base_fn + ".ceps"
np.save(data_fn, ceps)
print "Written ", data_fn
return data_fn
def generate_mfcc(voice, word, rate, path):
filename = path+"/ogg/{0}_{1}_{2}.ogg".format(word, voice, rate)
cmd = "say '{0}' -v{1} -r{2} -o '{3}'".format(word, voice, rate, filename)
os.system(cmd) # ogg aiff m4a or caff
signal, sample_rate = librosa.load(filename, mono=True)
# mel_features = librosa.feature.mfcc(signal, sample_rate)
# sample_rate, wave = scipy.io.wavfile.read(filename) # 2nd lib
mel_features, mspec, spec = mfcc(signal, fs=sample_rate, nceps=26)
# mel_features=python_speech_features.mfcc(signal, numcep=26, nfilt=26*2,samplerate=sample_rate) # 3rd lib
# print len(mel_features)
# print len(mel_features[0])
# print("---")
mel_features=np.swapaxes(mel_features,0,1)# timesteps x nFeatures -> nFeatures x timesteps
np.save(path + "/mfcc/%s_%s_%d.npy" % (word,voice,rate), mel_features)
def audio_stream(path, nfft): SR, X = wavfile.read(path) try: X = numpy.mean(X, axis=1) except: pass _, _, spec = mfcc(X, fs=SR, nfft=(nfft*2)) list_data = numpy.array(numpy.mean(spec, axis=0)[:nfft]/numpy.max(spec)) list_data = numpy.reshape(list_data, (len(list_data)/nfft, nfft)) return list_data
def extract_features(path):
sample_rate, X = scipy.io.wavfile.read(path)
fft_features = abs(scipy.fft(X) [:1000])
n = fft_features.size
timestep = (sample_rate/2.)/1000
max_time = timestep*n
freq = np.arange(0, max_time, timestep)
centroid = centroid = np.sum(fft_features*freq)/np.sum(fft_features)
ceps, mspec, spec = mfcc(X)
num_ceps = ceps.shape[0]
mfcc_features = np.mean(ceps[int(num_ceps*1/10):int(num_ceps*9/10)], axis = 0)
return fft_features, mfcc_features, centroid
def convert_to_mfcc(voice_path):
sample_rate, X = scipy.io.wavfile.read(voice_path)
ceps, mspec, spec = mfcc(X)
ave_cept = np.zeros((1, 13))
count = 0
for one_ceps in ceps:
if np.isnan(one_ceps[1]):
continue
ave_cept += one_ceps
count += 1
if count == 0:
return None
ave_cept /= count
return ave_cept
def record_plot_(self):
ax = self.figure.add_subplot(312)
ax.hold(False)
SR, X = wavfile.read('output.wav')
X = np.mean(X, axis=1)
plt.plot(X)
plt.xlim(0,len(X))
y = X + np.random.rand(X.shape[0])
ceps, mspec, spec = mfcc(y, fs=SR, nfft=nFFT)
ax = self.figure.add_subplot(313)
ax.hold(False)
plt.pcolormesh(mspec.T)
plt.xlim(0,len(mspec))
self.canvas.draw()
def add_mfcc(X, label_id, features, labels):
"""
Input
X: song data
label_id: label(genre) id
features: array of ffts
labels: array of labels
Description
extracts MFCC from X using scikits.talkbox and appends it to features.
"""
ceps, mspec, spec = mfcc(X)
num_ceps = len(ceps)
x = np.mean(ceps[int(num_ceps*1/10):int(num_ceps*9/10)], axis=0)
if np.isfinite(np.array(x).sum()):
features.append(x)
labels.append(label_id)
else:
raise ValueError('fft non-finite data')
