def DPGMM_test(cov_type, alpha_val):
#speakers_MFCC_dict = {}
#speaker_GMM_dict = {}
files = glob.glob(os.getcwd()+'\\speakers\\*.wav')
gauss_num = 32
iterator = 1
test_files = []
good = 0
bad = 0
total = 0
for file in files:
if file[-6:-4] == '09':
test_files.append(file)
for file in files:
#print(file)
if file[-6:-4] == '00': #file[len(file)-12:len(file)-9]
current_speaker = file[len(file)-10:len(file)-6]
#print("############# Calculate MFCC and DPGMM for ", current_speaker, " , speaker no ", str(iterator))
#if iterator == 572:
# print("Tu bedzie error")
merged_files = np.array([])
for i in range(0, 9):
current_file = wav.read(file[:-5]+str(i)+file[-4:])
merged_files = np.append(merged_files, current_file[1])
#print(type(merged_files))
speaker_MFCC = MFCC.extract(merged_files)
speaker_MFCC = speaker_MFCC[:, 1:]
#speakers_MFCC_dict[current_speaker] = speaker_MFCC
g = mixture.DPGMM(n_components=gauss_num, n_iter=100, covariance_type=cov_type, alpha=alpha_val)
g.fit(speaker_MFCC)
#speaker_model = np.array([g.means_, g.precs_, np.repeat(g.weights_[:, np.newaxis], 12, 1)])
#speaker_GMM_dict[current_speaker] = speaker_model
log_prob = -10000
winner = 'nobody'
for test_file in test_files:
current_test_speaker = test_file[len(test_file)-10:len(test_file)-6]
current_test_file = wav.read(test_file)
test_speaker_MFCC = MFCC.extract(current_test_file[1])
test_speaker_MFCC = test_speaker_MFCC[:, 1:]
temp_prob = np.mean(g.score(test_speaker_MFCC))
if temp_prob > log_prob:
log_prob = temp_prob
winner = current_test_speaker
if winner == current_speaker:
good += 1
else:
bad += 1
total +=1
#print(current_speaker, " speaker no ", str(iterator), " is similar to ", winner, " - log prob = ", str(log_prob))
#print("good = ", str(good), ", bad = ", str(bad), ", total = ", str(total))
iterator += 1
print("DPGMM (covariance_type - ", cov_type, ", alpha - ", str(alpha_val), "), Efficiency = ", str(good/total))
def mix_feature(tup):
mfcc = MFCC.extract(tup)
lpc = LPC.extract(tup)
if len(mfcc) == 0:
print >> sys.stderr, "ERROR.. failed to extract mfcc feature:", len(
tup[1])
return np.concatenate((mfcc, lpc), axis=1)
def select_events(nevents,nfeatures):
global groups
fftbins = 8192
featurewidth = 16
print "Selecting %d random spectral features.." % nfeatures
feature_bins = np.random.randint(featurewidth/2,(fftbins/8),nfeatures)
print "Selecting %d random audio events.." % nevents
events = np.random.randint(0,len(faudio)-grain_mid,nevents)
# Initialise features array with the first variable as index
features = np.zeros((14,nevents))
features[0] = np.arange(0,nevents)
print "Computing audio event spectrograms.."
# For each event..
for i in range(0,nevents):
# Calculate spectrogram for the event
_fftevent = faudio[events[i]:min(events[i]+1000,len(faudio))]*sig.hann(1000)
mfcc = MFCC.extract(_fftevent)
features[:,i] = np.append(i,mfcc)
#powerspec = abs(fft(_fftevent,fftbins)) ** 2
#melspec = np.dot(powerspec,melFilterBank(len(_fftevent)))
#logspec = np.log(melspec)
#mfcc = dct(logspec,type=2)
#print mfcc
# Calculate each feature for this event
#for j in range(0,nfeatures):
# features[j+1][i] = abs(np.mean(abs(mags[(feature_bins[j]-featurewidth/2):(feature_bins[j]+featurewidth/2)])))
print "Clustering events with K-Means algorithm.."
groups = kmeans(np.transpose(features),tracks,minit='points',iter=30)[1]
return [events,groups]
def Classify (self, sample, verbose = True):
length = len (sample)
features = MFCC.extract (numpy.frombuffer (sample, numpy.int16))
gestures = {}
for gesture in self.params:
d = []
for tsample in self.params[gesture]:
total_distance = 0
smpl_length = len(tsample)
if(numpy.abs(length - smpl_length) <= 0):
continue
for i in range (min (len (features), len (tsample))):
total_distance += dist.cityblock(features[i], tsample[i])
d.append (total_distance/float (i))
score = numpy.min(d)
gestures[gesture] = score
if(verbose):
print "Gesture %s: %f" % (gesture, score)
try:
if (score < minimum):
minimum = score
lowest = gesture
except:
minimum = score
lowest = gesture
if verbose:
print lowest, minimum
if(minimum < THRESHOLD):
return lowest
else:
return None
def test(filename, verbose = False):
rawdata = loadWAVfile(filename)
mfcc = MFCC.extract(rawdata, show=False)
#Test the hmm
HMM_Model.test(mfcc, verbose)
return
def GenerateParams (gestures, verbose = True):
params = {}
for gesture in gestures:
if(verbose):
print "Processing " + gesture
l = []
for sample in gestures[gesture]:
l.append (MFCC.extract (numpy.frombuffer (sample, numpy.int16)))
params[gesture] = l
return params
def train(filename, id):
rawdata = loadWAVfile(filename)
mfcc = MFCC.extract(rawdata, show=False)
model = VQ.Model(id)
#Train the VQ
model.train(mfcc)
#Train the HMM
create_file(mfcc, id)
return
def train(filename, id):
rawdata = loadWAVfile(filename)
mfcc = MFCC.extract(rawdata, show=False)
model = VQ.Model(id)
#Train the VQ
model.train(mfcc)
#Train the HMM
create_file(mfcc, id)
return
def load():
names = [
"Mathematics", "Biology", "PoliticalScience", "Statistics",
"Psychology"
]
sampledict = {}
for name in names:
sampledict[name] = []
for fname in glob.glob("Samples/" + name + " *"):
w = wread(fname)
sampledict[name].append(MFCC.extract(w[1])[:30])
return names, sampledict
def produce_mfcc(self, filename):
wav = wave.open(filename, "r")
x = np.fromstring(wav.readframes(self.sz), dtype=np.int16)
#(nchannels, sampwidth, framerate, nframes,
# comptype, compname) = wav.getparams()
mfcc = MFCC.extract(x)
match = self.lab_extractor.match(filename)
try:
label = match.group(1)
except:
label = "unknown"
print >> sys.stderr, "unknown labels encountered"
return (mfcc, label)
def add_to_database(url_, person_name_):
gmm_models = {}
if os.path.isfile('mfcc.mat'):
gmm_models = sio.loadmat('mfcc.mat')
print "Recording and processing...\n\n"
full_sound_model = read_radio_stream(url_)
wav.write('People\\'+person_name_+'.wav', 11025, full_sound_model/32767.0)
print "Calculating MFCC and saving the model..."
mfcc_features = MFCC.extract(full_sound_model)
mfcc_features = mfcc_features[:, 1:]
g = mixture.GMM(n_components=128)
g.fit(mfcc_features)
model = np.array([g.means_, g.covars_, np.repeat(g.weights_[:, np.newaxis], 12, 1)]) # weights have to be repeated to properly save the np array
print len(g.means_)
gmm_models[person_name_] = model
sio.savemat('mfcc_32.mat', gmm_models, oned_as='row')
def read_radio_stream(url_):
database = sio.loadmat('mfcc_16_fft256_GMM.mat')
database.pop('__header__')
database.pop('__version__')
database.pop('__globals__')
r2 = urllib.urlopen(url_)
pygame.mixer.init(44100, -16, 2, 2048)
print pygame.mixer.get_init()
chan1 = pygame.mixer.find_channel()
format = sound.AFMT_S16_LE
print sound.getODevices()
#snd_out = sound.Output(44100, 2, format)
dm = muxer.Demuxer('mp3')
dec = None
snd = None
print(r2.info())
print('###################\n')
#f = open('radio.mp3', 'wb')
#g = open('radio.wav', 'wb')
i = 0
while True: #i < 3:
samples = r2.read(15000)
frames = dm.parse(samples)
if dec is None:
# Open decoder
dec = acodec.Decoder(dm.streams[0])
#start = time.time()
sound_np_array = ansic_to_numpy(frames, dec)
#print (sound_np_array.shape[0])/44100.0
#elapsed = (time.time() - start)
#print 'decode and ndaray - %2.8f' %elapsed
#start = time.time()
to_play = np.array(np.repeat(sound_np_array[:, np.newaxis], 2, 1), dtype = 'int16')
sounds = pygame.sndarray.make_sound(to_play)
chan1.queue(sounds)
#elapsed = (time.time() - start)
#print 'to play - %2.8f' %elapsed
#start = time.time()
sound_np_array = decimate(sound_np_array, 4)
#elapsed = (time.time() - start)
#print 'downsample - %2.8f' %elapsed
#start = time.time()
mfcc_features = MFCC.extract(sound_np_array) #1.5s
mfcc_features = mfcc_features[:, 1:]
#elapsed = (time.time() - start)
#print 'mfcc - %2.8f' %elapsed
g = mixture.GMM(n_components=16)
log_prob = -10000
winner = 'nobody'
for key, values in database.iteritems():
try:
g.means_ = values[0, :, :]
g.covars_ = values[1, :, :]
g.weights_ = values[2, :, 1]
#start = time.time()
temp_prob = np.mean(g.score(mfcc_features))
#elapsed = (time.time() - start)
#print 'log-likelihood - %2.8f' %elapsed
if temp_prob > log_prob:
log_prob = temp_prob
winner = key
except TypeError:
print 'error dla ', key
print winner, log_prob
print('\n###################')
def get_mfcc_worker(fpath):
print('mfcc: ' + fpath)
fs, signal = wavfile.read(fpath)
mfcc = MFCC.extract(fs, signal)
return mfcc[:1500]
def collect(n=20):
obs = []
for i in xrange(n):
os.system("arecord -f S16_LE --rate=44100 -D hw:1,0 -d 3 test.wav")
obs.append(MFCC.extract(wavfile.read("test.wav")[1]))
return obs
def GMM_test(ii):
speakers_MFCC_dict = {}
speaker_GMM_dict = {}
files = glob.glob(os.getcwd()+'\\speakers\\*.wav')
gauss_num = 32
iterator = 1
num_iter = ii
if os.path.isfile('mfcc_'+str(gauss_num)+'.mat'):
speaker_GMM_dict = sio.loadmat('mfcc_'+str(gauss_num)+'.mat')
speaker_GMM_dict.pop('__header__')
speaker_GMM_dict.pop('__version__')
speaker_GMM_dict.pop('__globals__')
else:
for file in files:
#print(file)
if file[-6:-4] == '00': #file[len(file)-12:len(file)-9]
current_speaker = file[len(file)-10:len(file)-6]
print("############# Calculate MFCC and GMM for ", current_speaker, " , speaker no ", str(iterator))
#if iterator == 572:
# print("Tu bedzie error")
iterator += 1
merged_files = np.array([])
for i in range(0, 9):
current_file = wav.read(file[:-5]+str(i)+file[-4:])
merged_files = np.append(merged_files, current_file[1])
#print(type(merged_files))
speaker_MFCC = MFCC.extract(merged_files)
speaker_MFCC = speaker_MFCC[:, 1:]
speakers_MFCC_dict[current_speaker] = speaker_MFCC
g = mixture.GMM(n_components=gauss_num, n_iter=num_iter)
g.fit(speaker_MFCC)
speaker_model = np.array([g.means_, g.covars_, np.repeat(g.weights_[:, np.newaxis], 12, 1)])
speaker_GMM_dict[current_speaker] = speaker_model
sio.savemat('mfcc_'+str(gauss_num)+'.mat', speaker_GMM_dict, oned_as='row')
iterator = 1
good = 0
bad = 0
total = 0
for file in files:
if file[-6:-4] == '09':
g = mixture.GMM(n_components=gauss_num, n_iter=num_iter)
current_file = wav.read(file)
current_speaker = file[len(file)-10:len(file)-6]
#print(current_speaker, )
speaker_MFCC = MFCC.extract(current_file[1])
speaker_MFCC = speaker_MFCC[:, 1:]
log_prob = -10000
winner = 'nobody'
for key, values in speaker_GMM_dict.items():
try:
g.means_ = values[0, :, :]
g.covars_ = values[1, :, :]
g.weights_ = values[2, :, 1]
temp_prob = np.mean(g.score(speaker_MFCC))
if temp_prob > log_prob:
log_prob = temp_prob
winner = key
except TypeError:
print('error for ', key)
if current_speaker == winner:
good += 1
else:
bad += 1
total +=1
print(current_speaker, " speaker no ", str(iterator), " is similar to ", winner, " - log prob = ", str(log_prob))
print("good = ", str(good), ", bad = ", str(bad), ", total = ", str(total))
iterator += 1
print("GMM, n_iter = ", num_iter, ", Efficiency = ", str(good/total))
def get_mfcc_worker(fpath):
print('mfcc: ' + fpath)
fs, signal = wavfile.read(fpath)
mfcc = MFCC.extract(fs, signal)
return mfcc[:1500]
def CalculateMFCCs(self): # This function calculates and returns the MFCC from the given wavfile mfccs = MFCC.extract(self.wav_data) return mfccs
# python -i <name of this .py file> import numpy as np from scikits.audiolab import Sndfile SOUND_DIRECTORY = 'small_data_sample/right_whale' test_file = '%s/train12.aiff' % SOUND_DIRECTORY f = Sndfile(test_file, 'r') # Sndfile instances can be queried for the audio file meta-data fs = f.samplerate nc = f.channels enc = f.encoding # Reading is straightfoward data = f.read_frames(1000) # This reads the next 1000 frames, e.g. from 1000 to 2000, but as single precision data_float = f.read_frames(1000, dtype=np.float32) print data_float.shape import MFCC # data_float is a wave signal saved in a 1-D numpy array # mfcc is a 2-D numpy array, where each row is the # MFCC of a frame in data_float mfcc = MFCC.extract(data_float, show = True) # This will also plot the MFCC and the spectrogram # reconstructed from MFCC by inverse DCT
def collect(n=20):
obs = []
for i in xrange(n):
os.system("arecord -f S16_LE --rate=44100 -D hw:1,0 -d 3 test.wav")
obs.append(MFCC.extract(wavfile.read("test.wav")[1]))
return obs
