def load_data_shared(ind):
#Training and testing data
timit_data_train = genfromtxt('timit_data_1280_train.csv', delimiter=',')
timit_vwlname_train = genfromtxt('timit_vwlname_1280_train.csv', delimiter=',')
timit_vwlname_train[:] = [x - 1 for x in timit_vwlname_train]
timit_data_test = genfromtxt('timit_data_1280_test.csv', delimiter=',')
timit_vwlname_test = genfromtxt('timit_vwlname_1280_test.csv', delimiter=',')
timit_vwlname_test[:] = [x - 1 for x in timit_vwlname_test]
fs = 16000
datalen = 1280
narr = np.array([13, 26, 39]); #Number of features in each frame
i=0; j=0;
trainfeature=np.zeros((len(timit_data_train), (datalen*100/fs - 1)*narr[ind]))
for x in timit_data_train:
fbank_flat = logfbank(x,fs).flatten()
mfcc_flat = mfcc(x,fs).flatten()
if ind == 0:
trainfeature[i,:] = mfcc_flat
elif ind == 1:
trainfeature[i,:] = fbank_flat
else:
trainfeature[i,:] = np.concatenate((mfcc_flat, fbank_flat))
i = i+1
testfeature=np.zeros((len(timit_data_test), (datalen*100/fs - 1)*narr[ind]))
for x in timit_data_test:
fbank_flat = logfbank(x,fs).flatten()
mfcc_flat = mfcc(x,fs).flatten()
if ind == 0:
testfeature[j,:] = mfcc_flat
elif ind == 1:
testfeature[j,:] = fbank_flat
else:
testfeature[j,:] = np.concatenate((mfcc_flat, fbank_flat))
j = j+1
training_data = (trainfeature, timit_vwlname_train)
test_data = (testfeature, timit_vwlname_test)
# For now, I am using test data as validating data. Should change later.
validation_data = test_data
def shared(data):
"""Place the data into shared variables. This allows Theano to copy
the data to the GPU, if one is available.
"""
shared_x = theano.shared(
np.asarray(data[0], dtype=theano.config.floatX), borrow=True)
shared_y = theano.shared(
np.asarray(data[1], dtype=theano.config.floatX), borrow=True)
return shared_x, T.cast(shared_y, "int32")
return [shared(training_data), shared(validation_data), shared(test_data)]
def svm_baseline():
#### Change here
ind = 0; # 0 for mfcc, 1 for filterbank, 2 for both
narr = np.array([13, 26, 39]); # corresponding length of feature in a frame
#Training and testing data
timit_data_train = genfromtxt('timit_data_1280_train.csv', delimiter=',')
timit_vwlname_train = genfromtxt('timit_vwlname_1280_train.csv', delimiter=',')
timit_vwlname_train[:] = [x - 1 for x in timit_vwlname_train]
timit_data_test = genfromtxt('timit_data_1280_test.csv', delimiter=',')
timit_vwlname_test = genfromtxt('timit_vwlname_1280_test.csv', delimiter=',')
timit_vwlname_test[:] = [x - 1 for x in timit_vwlname_test]
fs = 16000
datalen = 1280
i=0; j=0;
trainfeature=np.zeros((len(timit_data_train), (datalen*100/fs - 1)*narr[ind]))
for x in timit_data_train:
fbank_flat = logfbank(x,fs).flatten()
mfcc_flat = mfcc(x,fs).flatten()
if ind == 0:
trainfeature[i,:] = mfcc_flat
elif ind == 1:
trainfeature[i,:] = fbank_flat
else:
trainfeature[i,:] = np.concatenate((mfcc_flat, fbank_flat))
i = i+1
testfeature=np.zeros((len(timit_data_test), (datalen*100/fs - 1)*narr[ind]))
for x in timit_data_test:
fbank_flat = logfbank(x,fs).flatten()
mfcc_flat = mfcc(x,fs).flatten()
if ind == 0:
testfeature[j,:] = mfcc_flat
elif ind == 1:
testfeature[j,:] = fbank_flat
else:
testfeature[j,:] = np.concatenate((mfcc_flat, fbank_flat))
j = j+1
training_data = (list(trainfeature), timit_vwlname_train)
test_data = (list(testfeature), timit_vwlname_test)
# train
clf = svm.SVC()
clf.fit(training_data[0], training_data[1])
# test
predictions = [int(a) for a in clf.predict(test_data[0])]
num_correct = sum(int(a == y) for a, y in zip(predictions, test_data[1]))
print "Using svm_baseline classifier:"
print "%s of %s values correct. %s percent " % (num_correct, len(test_data[1]),
(num_correct*100)/len(test_data[1]))
def getFeatures(signal, rate): """ Extracts Important Vocal Features author: chris """ if signal.shape[0] > mem_cut_off: mfcc,fbank = getFeatures(signal[mem_cut_off:], rate) return np.concatenate((fs.mfcc(signal[:mem_cut_off],rate),mfcc)), np.concatenate((fs.logfbank(signal,rate),fbank)) else: return fs.mfcc(signal,rate), fs.logfbank(signal,rate)
def getFeatures(signal, rate):
"""
Extracts Important Vocal Features
author: chris
"""
if signal.shape[0] > mem_cut_off:
mfcc, fbank = getFeatures(signal[mem_cut_off:], rate)
return np.concatenate((fs.mfcc(signal[:mem_cut_off],
rate), mfcc)), np.concatenate(
(fs.logfbank(signal, rate), fbank))
else:
return fs.mfcc(signal, rate), fs.logfbank(signal, rate)
def cal_bic(wavfilename, sadfilename):
sample_rate, wav = wavfile.read(wavfilename)
mfcc_feat = features.mfcc(wav, sample_rate)
ref = getsad_ref(sadfilename)
an_win_mfcc = mfcc_cut_vad_an_win(mfcc_feat, ref)
[time, bic_value] = bic(an_win_mfcc)
return time, bic_value
def fill(self, class_id):
"""Fills internal structer with new training samples.
Do not call directly.
:param class_id: class identification
"""
# get training samples
for i in range(len(self.sel_files)):
row = self.sel_files[0]
samples = self.all_files[row].samples
feat = mfcc(samples, 16000, winlen=0.030, appendEnergy=False, VAD=simpleVAD)
# add two symptoms from the middle
self.X.append(feat[int(len(feat) / 2 - 1)])
self.y.append(class_id)
self.X.append(feat[int(len(feat) / 2 + 1)])
self.y.append(class_id)
# clear from the list
del_iter = self.file_store.get_iter(Gtk.TreePath.new_from_indices([row]))
self.file_store.remove(del_iter)
del self.all_files[row]
# print results
if Classifier.new_training(self.X, self.y):
self.status_label.set_text("{0} vzorků, {1} tříd".format(len(self.X), len(self.class_names)))
else:
self.status_label.set_text("Klasifikátor potřebuje víc tříd")
def extract_mfcc_features(signal, win_len=0.0232, win_overlap=0.5, n_mel_bands=40, n_coefs=25, fs=48000, nfft=1024):
"""
Return feature vector for a one channel signal
Return same features as the one defined in the paper
Salamon, J., Jacoby, C., & Bello, J. (2014). A Dataset and Taxonomy for Urban Sound Research. ACM International Conference Onf Multimedia, (3). doi:10.1145/2647868.2655045
:param signal: one dimension array
:param win_len: length of window to split the signal into
:param win_overlap: overlap over window, 1 > win_overlap >= 0
:param n_mel_bands: numbers of mel bands to use
:param n_coefs: number of dct coefs to return
:param fs: signal sampling rate
:return: a dict of features array
"""
win_step = win_len * win_overlap # 50%
features = {}
res = mfcc(signal, samplerate=fs, winlen=win_len, winstep = win_step, nfilt = n_mel_bands, lowfreq=0,
highfreq = 22050, numcep=n_coefs, nfft=nfft) ## TODO revoir nfft.. je ne suis pas certain de comprendre a quoi ca correspond pour mel dans le papier il n'en parle pas.. surtour revoir si ca fonctionne avec nfft et fs... car bon
#print("fs {}, signal.shape {}".format(fs,signal.shape))
#print(res.shape)
features["minimum"] = np.min(res, axis=0)
features["maximum"] = np.max(res, axis=0)
features["median"] = np.median(res, axis=0)
features["mean"] = np.mean(res, axis=0)
features["variance"] = np.var(res, axis=0)
features["skewness"] = scipy.stats.skew(res, axis=0)
features["kurtosis"] = scipy.stats.kurtosis(res, axis=0)
features["mean_first_diff"] = np.mean(np.diff(res, axis=0), axis=0)
features["variance_first_diff"] = np.var(np.diff(res, axis=0), axis=0)
features["mean_second_diff"] = np.mean(np.diff(res, axis=0, n=2), axis=0)
features["var_second_diff"] = np.var(np.diff(res, axis=0, n=2), axis=0)
return features
def compute_mfcc(sig, rate, winlen=0.025, winstep=0.01, numcep = 12, nfilt=26, nfft=512, lowfreq=0, highfreq=None, preemph=0.97, ceplifter=22, include_energy=True, snip_edges = True): if snip_edges: #snip the edges sig = snip(sig, rate, winlen, winstep) return mfcc(sig, rate, winlen, winstep, numcep, nfilt, nfft, lowfreq, highfreq, preemph, ceplifter, include_energy)
def features_from_base(basepath, order=0):
(females, males) = read_speakers(basepath)
# list of list (sorted)
female_utterances_list = [
read_utterances(basepath, female) for female in females
]
male_utterances_list = [read_utterances(basepath, male) for male in males]
# utterances as Wave objects
female_utterances_list = read_utterances_files(basepath,
female_utterances_list, 'f')
male_utterances_list = read_utterances_files(basepath,
male_utterances_list, 'm')
for utterances in female_utterances_list:
for utterance in utterances:
uttMFCCs = features.mfcc(utterance.signal,
samplerate=utterance.sample_rate,
numcep=19,
highfreq=utterance.sample_rate / 2)
if (order > 0):
uttMFCCs = features.appendDeltasAllFrames(uttMFCCs, order)
print(uttMFCCs.shape)
print(uttMFCCs)
print()
def run_tests(test_files):
# Classify input data
for test_file in test_files:
# Read input file
sampling_freq, signal = wavfile.read(test_file)
# Extract MFCC features
with warnings.catch_warnings():
warnings.simplefilter('ignore')
features_mfcc = mfcc(signal, sampling_freq)
# Define variables
max_score = -float('inf')
output_label = None
# Run the current feature vector through all the HMM
# models and pick the one with the highest score
for item in speech_models:
model, label = item
score = model.compute_score(features_mfcc)
if score > max_score:
max_score = score
predicted_label = label
# Print the predicted output
start_index = test_file.find('/') + 1
end_index = test_file.rfind('/')
original_label = test_file[start_index:end_index]
print('\nOriginal: ', original_label)
print('Predicted:', predicted_label)
def vector_quantize(myfiles, outdir, model): #given a list of files transform them to spectral vectors and compute the KMeans VQ
for f in myfiles:
print "Quantizing: ", f
(rate, sig) = wav.read(f)
#print rate, sig.shape
#get the spectral vectors
mfcc_feat = mfcc(sig,rate)
#print mfcc_feat.shape
fbank_feat = mfcc_feat #logfbank(sig,rate) #this has the spectral vectors now
#print fbank_feat.shape
val = model.predict(fbank_feat)
#fcomps = os.path.split(f) #file components path, filename
fcomps = f.split("/")
fn = fcomps[-2]+"/"+fcomps[-1].split('.')[0] + '_vq.txt'
#outpath = os.path.join(fcomps[0], 'outputs')
fn = os.path.join(outdir, fn)
d = os.path.dirname(fn)
if not os.path.exists(d):
os.makedirs(d)
#print fn
#val = trim_background(val)
#raw_input("enter...")
f = open(fn, 'wb')
for v in val:
f.write(str(v) + '\n')
f.close()
print 'output vector quantized file: ', f, ' written'
return
def training():
'''
Takes input signal and searches current dataset for hit.
If hit, then add to correct dataset.
If miss, asks user for currect input and adds to dataset.
'''
print("please speak a word into the microphone")
record_to_file('training.wav')
print("done - result written to training.wav")
(rate, sig) = wav.read("training.wav")
mfcc_feat = mfcc(sig, rate)
fbank_feat = logfbank(sig, rate)
recording = fbank_feat[1:3, :]
testing = check_for_match(recording)
verify = raw_input("did you say " + testing + " ")
if verify is 'y':
parse_array(recording, testing)
if verify is 'n':
correct_word = input("what word did you mean? ")
print correct_word
parse_array(recording, correct_word)
def generate_testing_mfccs(myfiles, outdir):
for f in myfiles:
print "Generating MFCCs for: ", f
(rate, sig) = wav.read(f)
#print rate, sig.shape
#get the spectral vectors
mfcc_feat = mfcc(sig,rate)
#mfcc_feat = scaler.transform(mfcc_feat)
#fcomps = os.path.split(f) #file components path, filename
fcomps = f.split("/")
fn = fcomps[-2]+"/"+fcomps[-1].split('.')[0] + '_mfcc.txt'
#outpath = os.path.join(fcomps[0], 'outputs')
fn = os.path.join(outdir, fn)
d = os.path.dirname(fn)
if not os.path.exists(d):
os.makedirs(d)
f = open(fn, 'wb')
final_mfccs_str=""
for vector in mfcc_feat:
str_mfcc = ""
for element in vector:
str_mfcc +=str(element)+","
str_mfcc = str_mfcc[:-1]
final_mfccs_str += str_mfcc+"\n"
f.write(final_mfccs_str)
f.close()
print 'output MFCC file: ', f, ' written'
return
def __init__(self, filename):
self.filename = filename
self.frequency, self.sound = wavfile.read(wavFilesPath + filename)
self.channel1 = self.sound[:, 0]
self.channel2 = self.sound[:, 1]
self.duration = len(self.sound) / self.frequency
self.mfccFeatures = mfcc(self.sound, self.frequency)
def extractLow(signal):
return mfcc(signal,
samplerate=SAMPLING_RATE,
winlen=LO_FRAME_DURATION,
winstep=LO_FRAME_STEP,
numcep=NUM_CEPTRUM,
appendEnergy=True)
def get_data(rootdir = TIMIT_main_dir):
inputs = []
targets = []
for dir_path, sub_dirs, files in os.walk(rootdir):
for file in files:
if (os.path.join(dir_path, file)).endswith('.wav'):
wav_file_name = os.path.join(dir_path, file)
input_data, f_s = sf.read(wav_file_name)
# mfcc_feat = MFCC_input(mfcc(input_data,f_s))
mfcc_feat = mfcc(input_data,f_s)
#Delta features
delta_feat = mfcc_feat[:-1]-mfcc_feat[1:]
#Delta-Delta features
deltadelta_feat = delta_feat[:-1]-delta_feat[1:]
#Removing the first two frames
mfcc_feat = mfcc_feat[2:]
delta_feat = delta_feat[1:]
#Concatenating mfcc, delta and delta-delta features
full_input = np.concatenate((mfcc_feat,delta_feat,deltadelta_feat), axis=1)
inputs.append(np.asarray(full_input, dtype=theano.config.floatX))#Rakeshvar wants one frame along each column but i am using Lasagne
text_file_name = wav_file_name[:-4] + '.txt'
target_data_file = open(text_file_name)
target_data = str(target_data_file.read()).lower().translate(None, '!:,".;?')
# target_data = str(target_data_file.read()).lower().translate(str.maketrans('','', '!:,".;?'))
target_data = target_data[8:-1]#No '.' in lexfree dictionary
targets.append(target_data)
return inputs, targets
def shifted_delta_cepstra(self, wav_fn, delta=1, shift=3, k_conc=3):
"""
:param
delta: represents the time advance and delay for the sdc
k_conc: is the number of blocks whose delta coefficients are concd
shift: is the time shift between consecutive blocks
Shifted delta cepstra are feature vectors created by concatenating
delta cepstra computed across multiple speech frames.
See the paper
PA Torres-Carrasquillo et al (2002)
Approaches to language identification using
Gaussian mixture models and Shifted delta cepstral features.
"""
(rate, sig) = wav.read(wav_fn)
mfcc_feats = mfcc(sig, rate)
# len(mfcc) == 39 == 3 * (12 cepstral + 1 energy)
# TODO include original cepstra as well?
delta_feats = mfcc_feats[delta:] - mfcc_feats[:-delta]
output_duration = delta_feats.shape[0] - shift * k_conc
shifted = np.zeros(
(output_duration, (k_conc + 1) * delta_feats.shape[1]))
mfcc_dim = mfcc_feats.shape[1]
shifted[:, 0:mfcc_dim] = mfcc_feats[:output_duration]
for i in xrange(output_duration):
shifted[i,
mfcc_dim:] = delta_feats[i:i +
k_conc * shift:shift, :].reshape(
(1, -1))
logger.debug('{} --> {}'.format(mfcc_feats.shape, shifted.shape))
return shifted
def build_codebook(
trgfile,
codesize=32,
fname=None
): # given a training file constructs the codebook using kmeans
#print "Codesize is ", codesize
(rate, sig) = wav.read(trgfile)
print rate, sig.shape
#get the spectral vectors
print("MFCC generation begins")
mfcc_feat = mfcc(sig, rate)
print("MFCC generation ends")
print mfcc_feat.shape
sys.exit(0)
#print("Fbank creation begins")
#fbank_feat = logfbank(sig,rate) #this has the spectral vectors now
#print("Fbank creation ends")
#print fbank_feat.shape
print "codesize = ", codesize
km = KMeans(n_clusters=codesize)
#km.fit(fbank_feat)
km.fit(mfcc_feat)
if fname != None:
pickle.dump(km, open(fname, 'wb'))
return km
def extractLow(signal):
return mfcc(signal,
samplerate = SAMPLING_RATE,
winlen = LO_FRAME_DURATION,
winstep = LO_FRAME_STEP,
numcep = NUM_CEPTRUM,
appendEnergy = True)
def main():
if len(sys.argv) < 2:
sys.stderr.write('Usage: python ' + sys.argv[0] + ' lang_test.wav')
sys.exit(1)
file = sys.argv[1]
languages = pickle.load( open('languages.dat', 'r') )
(rate,sig) = wav.read(file) # returns (sample rate, numpy.ndarray of samples)
mfcc_feat = mfcc(sig,rate)
mfccs_deltas = recognizer_util.get_deltas(mfcc_feat, 0)
mfccs_deltas_ddeltas = recognizer_util.get_deltas(mfccs_deltas, 13)
test_avg = recognizer_util.col_avg(mfccs_deltas_ddeltas)
results = {}
for language in languages.keys():
dist = get_distance(languages[language], test_avg)
results[dist] = language
sorted = results.keys()
sorted.sort()
print
language = results[sorted[0]]
sys.stdout.write(language)
print
def shifted_delta_cepstra(self, wav_fn, delta=1, shift=3, k_conc=3):
"""
:param
delta: represents the time advance and delay for the sdc
k_conc: is the number of blocks whose delta coefficients are concd
shift: is the time shift between consecutive blocks
Shifted delta cepstra are feature vectors created by concatenating
delta cepstra computed across multiple speech frames.
See the paper
PA Torres-Carrasquillo et al (2002)
Approaches to language identification using
Gaussian mixture models and Shifted delta cepstral features.
"""
(rate,sig) = wav.read(wav_fn)
mfcc_feats = mfcc(sig,rate)
# len(mfcc) == 39 == 3 * (12 cepstral + 1 energy)
# TODO include original cepstra as well?
delta_feats = mfcc_feats[delta:] - mfcc_feats[:-delta]
output_duration = delta_feats.shape[0] - shift*k_conc
shifted = np.zeros((output_duration,
(k_conc + 1) * delta_feats.shape[1]))
mfcc_dim = mfcc_feats.shape[1]
shifted[:,0:mfcc_dim] = mfcc_feats[:output_duration]
for i in xrange(output_duration):
shifted[i,mfcc_dim:] = delta_feats[i:i+k_conc*shift:shift,
:].reshape((1,-1))
logger.debug('{} --> {}'.format(mfcc_feats.shape, shifted.shape))
return shifted
def _gen_features(self, data_dir, outfile):
""" Generates a csv file containing labeled lines for each speaker """
with open(outfile, 'w') as ohandle:
melwriter = csv.writer(ohandle)
speakers = os.listdir(data_dir)
for spkr_dir in speakers:
for soundclip in os.listdir(os.path.join(data_dir, spkr_dir)):
# generate mel coefficients for the current clip
clip_path = os.path.abspath(
os.path.join(data_dir, spkr_dir, soundclip))
sample_rate, data = wavfile.read(clip_path)
ceps = mfcc(data, sample_rate)
# write an entry into the training file for the current speaker
# the vector to store in csv file contains the speaker's name at the end
fvec = self._mfcc_to_fvec(ceps)
fvec.append(spkr_dir)
logging.debug(
fvec) # see the numbers [as if they make sense ]
# write one row to the csv file
melwriter.writerow(fvec)
def create_mfcc(method, filename):
"""Perform standard preprocessing, as described by Alex Graves (2012)
http://www.cs.toronto.edu/~graves/preprint.pdf
Output consists of 12 MFCC and 1 energy, as well as the first derivative of these.
[1 energy, 12 MFCC, 1 diff(energy), 12 diff(MFCC)
method is a dummy input!!"""
(rate, sample) = wav.read(filename)
mfcc = features.mfcc(sample,
rate,
winlen=0.025,
winstep=0.01,
numcep=13,
nfilt=26,
preemph=0.97,
appendEnergy=True)
derivative = np.zeros(mfcc.shape)
for i in range(1, mfcc.shape[0] - 1):
derivative[i, :] = mfcc[i + 1, :] - mfcc[i - 1, :]
out = np.concatenate((mfcc, derivative), axis=1)
return out, out.shape[0]
def compute_features(filename):
fs,audio_array = wav.read(filename)
mfcc_25 = mfcc(audio_array,samplerate=fs,winlen=0.064,winstep=0.032,numcep=25,
nfilt=40,nfft=512,lowfreq=0,highfreq=fs/2,preemph=0,
ceplifter=0,appendEnergy=True)
first = np.diff(mfcc_25, axis=0)
second = np.diff(first, axis=0)
minimum = np.amin(mfcc_25,axis=0)
maximum = np.amax(mfcc_25,axis=0)
median = np.median(mfcc_25,axis=0)
mean = np.mean(mfcc_25,axis=0)
variance = np.var(mfcc_25,axis=0)
skewness = scipy.stats.skew(mfcc_25,axis=0)
kurtosis = scipy.stats.kurtosis(mfcc_25,axis=0)
first_mean = np.mean(first,axis=0)
first_variance = variance = np.var(first,axis=0)
second_mean = np.mean(second,axis=0)
second_variance = variance = np.var(second,axis=0)
features = np.concatenate((minimum, maximum, median, mean, variance, skewness, kurtosis, first_mean, first_variance, second_mean, second_variance), axis=0)
return features
def vector_quantize(
myfiles, outdir, model
): #given a list of files transform them to spectral vectors and compute the KMeans VQ
for f in myfiles:
print "Quantizing: ", f
(rate, sig) = wav.read(f)
#print rate, sig.shape
#get the spectral vectors
mfcc_feat = mfcc(sig, rate)
#print mfcc_feat.shape
fbank_feat = mfcc_feat #logfbank(sig,rate) #this has the spectral vectors now
#print fbank_feat.shape
val = model.predict(fbank_feat)
#fcomps = os.path.split(f) #file components path, filename
fcomps = f.split("/")
fn = fcomps[-2] + "/" + fcomps[-1].split('.')[0] + '_vq.txt'
#outpath = os.path.join(fcomps[0], 'outputs')
fn = os.path.join(outdir, fn)
d = os.path.dirname(fn)
if not os.path.exists(d):
os.makedirs(d)
#print fn
val = trim_background(val)
#raw_input("enter...")
f = open(fn, 'wb')
for v in val:
f.write(str(v) + '\n')
f.close()
print 'output vector quantized file: ', f, ' written'
return
def feature_extract_mfcc(self, sound, rate):
"""
extract every features for training
- frequency space: MFCC. pitch
:return:
"""
reg = re.compile('(\d+)-(\d+)-(\d+).wav')
#plotter.plot_frame(sound, show=True)
mfcc0 = mfcc(sound.reshape(1, -1),
rate,
winlen=cfg.frame,
winstep=cfg.step,
nfft=1536,
winfunc=np.hamming)
mfcc0 = mfcc0 - np.mean(mfcc0)
mfcc1 = delta(mfcc0, 3)
mfcc2 = delta(mfcc1, 3)
mfcc0 = scale(mfcc0)
'''
if audio in ['01','00']:
print(filename)
plotter.plot_mfcc(mfcc0,'311')
plotter.plot_mfcc(mfcc1,'312')
plotter.plot_mfcc(mfcc2,'313')
plotter.show()
'''
return (mfcc0, mfcc1, mfcc2), min(len(mfcc0), 200)
def compare(control_path, exp_path):
"""
Compares two wav files and returns a score. Uses mel frequency ceptrum coefficients as well as dynamic time warping.
:param control_path: the 'correct' wav - what you are comparing to
:param exp_path: the unknown wav
"""
(rate, sig) = wavread(control_path)
(rate2, sig2) = wavread(exp_path)
x = mfcc(sig, rate)
y = mfcc(sig2, rate2)
dist, cost, acc = dtw.dtw(x, y, dist=lambda x, y: dtw.norm(x - y, ord=1))\
return dist
def compare(control_path, exp_path):
"""
Compares two wav files and returns a score. Uses mel frequency ceptrum coefficients as well as dynamic time warping.
:param control_path: the 'correct' wav - what you are comparing to
:param exp_path: the unknown wav
"""
(rate,sig) = wavread(control_path)
(rate2,sig2) = wavread(exp_path)
x = mfcc(sig,rate)
y = mfcc(sig2,rate2)
dist, cost, acc = dtw.dtw(x, y, dist=lambda x, y: dtw.norm(x - y, ord=1))\
return dist
def get_data(rootdir=TIMIT_main_dir):
inputs = []
targets = []
for dir_path, sub_dirs, files in os.walk(rootdir):
for file in files:
if (os.path.join(dir_path, file)).endswith('.wav'):
wav_file_name = os.path.join(dir_path, file)
input_data, f_s = sf.read(wav_file_name)
# mfcc_feat = MFCC_input(mfcc(input_data,f_s))
mfcc_feat = mfcc(input_data, f_s)
#Delta features
delta_feat = mfcc_feat[:-1] - mfcc_feat[1:]
#Delta-Delta features
deltadelta_feat = delta_feat[:-1] - delta_feat[1:]
#Removing the first two frames
mfcc_feat = mfcc_feat[2:]
delta_feat = delta_feat[1:]
#Concatenating mfcc, delta and delta-delta features
full_input = np.concatenate(
(mfcc_feat, delta_feat, deltadelta_feat), axis=1)
inputs.append(
np.asarray(full_input, dtype=theano.config.floatX)
) #Rakeshvar wants one frame along each column but i am using Lasagne
text_file_name = wav_file_name[:-4] + '.txt'
target_data_file = open(text_file_name)
target_data = str(target_data_file.read()).lower().translate(
None, '!:,".;?')
# target_data = str(target_data_file.read()).lower().translate(str.maketrans('','', '!:,".;?'))
target_data = target_data[8:-1] #No '.' in lexfree dictionary
targets.append(target_data)
return inputs, targets
def read_data(files_amount, total_length, nc = 13, path=''): for i in range(files_amount): (rate,sig) = wav.read(path + str(i) + '.wav') mfcc_feat = mfcc(sig,rate, numcep = nc) mfcc_feat = np.reshape(mfcc_feat, (len(mfcc_feat)*nc, 1)) if any(np.isnan(mfcc_feat)) or any(np.isinf(mfcc_feat)): ind = [x for x in range(len(mfcc_feat)) if np.isnan(mfcc_feat[x]) or np.isinf(mfcc_feat[x])] for x in ind: mfcc_feat[x] = 0 if i == 0: mfcc_data = mfcc_feat if i != 0: if len(mfcc_feat) == total_length: mfcc_data = np.hstack((mfcc_data, mfcc_feat)) else: if len(mfcc_feat) > total_length: mfcc_data = np.hstack((mfcc_data, mfcc_feat[:(total_length)])) else: xx = np.vstack((mfcc_feat,np.reshape(np.asarray([0] * (total_length-len(mfcc_feat))), (total_length-len(mfcc_feat),1) ))) mfcc_data = np.hstack((mfcc_data, xx)) return mfcc_data
def crossover(playlist_1, playlist_2, playlist_size): #Crosses over playlists global all_playlistsl #print "Playlist_size: ", playlist_size one = all_playlists[playlist_1] two = all_playlists[playlist_2] child = Playlist(playlist_size) one_percentage = (one.fitness / float(one.fitness + two.fitness)) #print "One %: ", one_percentage one_genes = int(floor(playlist_size * one_percentage)) print "One genes: ", one_genes one_copy = copy.deepcopy(one) two_copy = copy.deepcopy(two) #Get genes from first parent for i in range(one_genes): #if len(one_copy.songs) <= 1: # two_genes+=1 # break all_songs.append(file) woteva = mfcc(sig, rate) woteva = reduce_matrix(woteva, 500) return woteva
def crossover(playlist_1, playlist_2, playlist_size):
#Crosses over playlists
global all_playlistsl
#print "Playlist_size: ", playlist_size
one = all_playlists[playlist_1]
two = all_playlists[playlist_2]
child = Playlist(playlist_size)
one_percentage = (one.fitness / float(one.fitness + two.fitness))
#print "One %: ", one_percentage
one_genes = int(floor(playlist_size * one_percentage))
print "One genes: ", one_genes
one_copy = copy.deepcopy(one)
two_copy = copy.deepcopy(two)
#Get genes from first parent
for i in range(one_genes):
#if len(one_copy.songs) <= 1:
# two_genes+=1
# break
all_songs.append(file)
woteva = mfcc(sig, rate)
woteva = reduce_matrix(woteva, 500)
return woteva
def make_mean_mfcc(filename):
try:
(rate, sig) = wav.read(filename)
mfcc_feat = mfcc(sig, rate)
avg_mfcc = np.mean(mfcc_feat, axis = 0)
return avg_mfcc
except:
pass
def predict(self, signal, fs = 44100):
if len(signal.shape) > 1:
signal = signal[:, 0]
signal_new = remove_silence(fs, signal)
# if len(signal_new) < len(signal) / 4:
# return "Silence"
mfcc_vecs = mfcc(signal_new, fs, numcep = 15)
return self.predict_feat(mfcc_vecs)
def readSegFeat(start_t, end_t, signal, sr):
try:
sig = signal[int(sr * start_t):int(sr * end_t)]
except:
sig = signal[int(sr * start_t):-1]
cleansig = remove_silence(sr, sig)
mfcc_vecs = mfcc(cleansig, sr, numcep=15)
return mfcc_vecs
def extract_feats(signal, sr):
feats = mfcc(signal, sr)
#fbank_feat = logfbank(signal, sr, nfilt = 17)
#feats = np.hstack((mfcc_feat, fbank_feat))
mu = np.mean(feats, axis = 0)
sigma = np.std(feats, axis = 0)
feature = (feats - mu) / sigma
return feature
def predict(self, signal, fs=44100):
if len(signal.shape) > 1:
signal = signal[:, 0]
signal_new = remove_silence(fs, signal)
# if len(signal_new) < len(signal) / 4:
# return "Silence"
mfcc_vecs = mfcc(signal_new, fs, numcep=15)
return self.predict_feat(mfcc_vecs)
def readSegFeat(start_t, end_t, signal, sr):
try:
sig = signal[int(sr * start_t) : int(sr * end_t)]
except:
sig = signal[int(sr * start_t) : -1]
cleansig = remove_silence(sr, sig)
mfcc_vecs = mfcc(cleansig, sr, numcep = 15)
return mfcc_vecs
def addMFCC(data_dict):
for name in data_dict:
data_dict[name]['mfcc'] =[]
audio_path = data_dict[name]['raw']
sr, WAV = wav.read(audio_path)
MFCC = mfcc(WAV, sr)
data_dict[name]['mfcc'].append(MFCC)
return data_dict
def extract_mfcc():
#nadi sve .wav fileove u audio direktoriju
audioDir = "audio/"
if not os.path.exists(audioDir):
os.makedirs(audioDir)
print "Ne postoji audio direktorij!"
audioFiles = []
for file in os.listdir(audioDir):
if fnmatch.fnmatch(file, '*.wav'):
audioFiles.append(file)
#ispis broja pronadenih .wav datoteka u audio direktoriju
print ""
print "Pronasao sam %d audio zapisa u %s direktoriju!" % (len(audioFiles),
audioDir)
print ""
#petlja koja prolazi kroz sve .wav datoteke unutar audio direktorija
for x in range(0, len(audioFiles)):
#ime .wav audio zapisa, ime .png grafa, ime .txt formata audio zapisa, ime direktorija za pohranu svega
filename = audioFiles[x]
floatFile = filename.split(".")[0] + ".txt"
directory = audioDir + filename.split(".")[0]
#provjere dali postoji datoteka ili direktoriji
if not os.path.isfile(audioDir + filename):
sys.exit("File does not exist!")
if not os.path.exists(directory):
os.makedirs(directory)
if os.path.exists(directory):
shutil.rmtree(directory)
os.makedirs(directory)
#citanje wav datoteke i slanje za izracun mfcc
(rate, sig) = wav.read(audioDir + filename)
mfcc_feat = mfcc(sig,
rate,
winlen=0.025,
winstep=0.01,
numcep=13,
preemph=0.99)
#ispis imena izracunatih .wav datoteka
print "MFCC karakteristike izracunate za %s!" % filename
#spremanje signala u float formatu i mfcc znacajki u .txt datoteke
np.savetxt(directory + "/" + floatFile, sig, fmt="%.4f")
np.savetxt(directory + "/mfcc_features.txt",
mfcc_feat,
fmt="%.16f",
delimiter=",")
#ispis broja .wav datoteka iz kojih smo izvukli MFCC
print "Ukupno %d MFCC karakteristika izracunato!" % len(audioFiles)
return 1
def feature_extract(wav_name, winlen=0.025, winstep=0.01):
"""This function returns (mfcc) feature vectors extracted from wav_name"""
rate, signal = wav.read(wav_name)
signal = numpy.sum(signal, axis=1) / signal.shape[1]
signal = sigproc.framesig(signal, rate * winlen, rate * winstep)
signal = vad.vad_filter(signal)
signal = sigproc.deframesig(signal, 0, rate * winlen, rate * winstep)
mfcc_feat = mfcc(signal, rate)
return mfcc_feat
def source_save(self):
for key, value in self.train_file_map.items():
wav = wave.open(key)
rate = wav.getframerate()
#Binary file needs to be munged because it is 2 channel
#encoding is 24bit signed integer Pulse Code Modulation (PCM)
#with a 44.1kHz sampling
#The initial way I was hoping to munge the raw byte data
#nframes = wav.getnframes()
#buf = wav.readframes(nframes)
#data is 24 bits in 3 bytes. np.int24 does not exist!
#dt = np.dtype(np.int24)
#data is in little endian format
#dt = dt.newbyteorder('<')
#sig = np.frombuffer(buf,dtype=dt)
#numpy doesn't support int24 yet so had to use this:
#http://stackoverflow.com/questions/12080279/how-do-i-create-a-numpy-dtype-that-includes-24-bit-integers
rawdatamap = np.memmap(key, dtype=np.dtype('u1'), mode='r')
usablebytes = rawdatamap.shape[0] - rawdatamap.shape[0] % 12
frames = int(usablebytes / 12)
rawbytes = rawdatamap[:usablebytes]
#This line is the difficult part which required stackoverflow, it makes the data into 32bit data,
#but because it is actually 24bit data there is included redundant data in the first byte.
realdata = as_strided(rawbytes.view(np.int32),
strides=(
12,
3,
),
shape=(frames, 2))
#This ANDs the bits by a byte mask of the last 24bits, to get rid of the redundant data
sig = realdata & 0x00ffffff
#mfcc is mel frequency cepstral coefficent
#http://practicalcryptography.com/miscellaneous/machine-learning/guide-mel-frequency-cepstral-coefficients-mfccs/
#mfcc_feat needs to be stored in MongoDB, it is a numpy array that is 5999 in length
#Each Audio file is a scene which is being classified, one of feature vectors
#used to classfy the scene is
#the mfcc_feat array
#mfcc will return an array of that is 5999 rows by 13 columns
#Each column is a feature vector for training the classifier for that audio sample's
#class (i.e. tram, park)
#The window length for analysis is 0.025 seconds,
#the window step between windows is 0.01 seconds.
#This is the entire array of feature vectors for each audio sample.
#Additional feature vectors might be added later but this is good for inital tests.
mfcc_feat = mfcc(sig, samplerate=rate)
#Insert records into mongodb
self.insert_mongo(self.mfcc_fv, mfcc_feat, key, value)
def _extract_mfcc(filename):
"""Extracts mfccs from wav files"""
savename = filename[0:len(filename) - 4] + '.mfc'
samp_rate, X = read(filename)
# ceps, mspec, spec = mfcc(X)
ceps = feat.mfcc(X, samp_rate)
num_ceps = ceps.shape[0]
x = np.mean(ceps[int(num_ceps * 1 / 10):int(num_ceps * 9 / 10)], axis=0)
np.save(savename, x)
def start(seed):
#Prepare global variables and the seed_mfcc
(rate, sig) = wav.read(seed)
for file in glob.glob("*.wav"):
all_songs.append(file)
woteva = mfcc(sig, rate)
woteva = reduce_matrix(woteva, 500)
return woteva
def learn(wav_filename, old_data=None):
rate, signal = wav.read(wav_filename, 'r')
mfcc_feat = mfcc(signal, rate)
if not old_data == None:
mfcc_feat = np.concatenate((mfcc_feat, old_data))
gmm = mixture.GMM(GMM_CLUSTERS)
gmm.fit(mfcc_feat)
return gmm, mfcc_feat
def plot_bic(wavfilename, sadfilename):
sample_rate, wav = wavfile.read(wavfilename)
mfcc_feat = features.mfcc(wav, sample_rate)
ref = getsad_ref(sadfilename)
an_win_mfcc = mfcc_cut_vad_an_win(mfcc_feat, ref)
[time, bic_value] = bic(an_win_mfcc)
pyplot.plot(time, bic_value)
pyplot.scatter(time, bic_value)
pyplot.show()
def start(seed):
#Prepare global variables and the seed_mfcc
(rate, sig) = wav.read(seed)
for file in glob.glob("*.wav"):
all_songs.append(file)
woteva = mfcc(sig, rate)
woteva = reduce_matrix(woteva, 500)
return woteva
def generate_speech(addr):
try:
(rate,sig) = wav.read(addr)
#plot_graf(sig, rate)
mfcc_feat = mfcc(sig,rate, highfreq=4000, numcep=20)
return lbg.generate_codebook(mfcc_feat, 16)[0]
except ValueError:
print("ValueErroe: Not a WAV file \nExit.")
return -1
def get_binned_features(self):
binned_features = {}
for i in range(len(self.binned_signals)):
binned_features[i] = features.mfcc(self.binned_signals[i],
self.sample_rate,
winlen=self.feature_winlen,
numcep=self.feature_numcep,
nfilt=self.feature_nfilt)
return binned_features
def get_MFCC_feature(sig, rate):
p_array = mfcc(sig, rate, winlen=0.025, winstep=0.01) # 获取梅尔倒谱系数
col_num = p_array.shape[1]
feature_array = []
for ii in range(col_num):
# test1. 取某一维的标准差
# feature_array.append(np.std(p_array[:,ii]))
feature_array.append(p_array[:, ii])
return feature_array
def feature_extract(wav_name, winlen=0.025, winstep=0.01):
"""This function returns (mfcc) feature vectors extracted from wav_name"""
rate, signal = wav.read(wav_name)
signal = numpy.sum(signal, axis=1)/signal.shape[1]
signal = sigproc.framesig(signal, rate*winlen, rate*winstep)
signal = vad.vad_filter(signal)
signal = sigproc.deframesig(signal, 0, rate*winlen, rate*winstep)
mfcc_feat = mfcc(signal, rate)
return mfcc_feat
def MFCC(data, samp):
mfcc_feat = mfcc(data, samp)
mMin = mfcc_feat.min()
mMax = mfcc_feat.max()
mfcc_feat -= mMin
mfcc_feat *= 255 / mfcc_feat.max()
outImg = np.array(mfcc_feat, np.uint8)
return outImg
def generate(self,testsample):
(rate,audio) = wav.read(testsample.path)
# grab first channel
one_channel = _extract_single_channel(audio)
N = len(audio)
mfcc_feat = mfcc(one_channel,rate)
cols=mfcc_feat.shape[0]*mfcc_feat.shape[1]
return mfcc_feat.reshape((1,cols))[0]
def MFCC(data, samp):
mfcc_feat = mfcc(data,samp)
mMin = mfcc_feat.min()
mMax = mfcc_feat.max()
mfcc_feat -= mMin
mfcc_feat *= 255/mfcc_feat.max()
outImg = np.array(mfcc_feat, np.uint8)
return outImg
def build_models(input_folder):
# Initialize the variable to store all the models
speech_models = []
# Parse the input directory
for dirname in os.listdir(input_folder):
# Get the name of the subfolder
subfolder = os.path.join(input_folder, dirname)
if not os.path.isdir(subfolder):
continue
# Extract the label
label = subfolder[subfolder.rfind('/') + 1:]
# Initialize the variable to store the training data
X = np.array([])
# Create a list of files to be used for training
# We will leave one file per folder for testing
training_files = [
x for x in os.listdir(subfolder) if x.endswith('.wav')
][:-1]
# Iterate through the training files and build the models
for filename in training_files:
# Extract the current file path
filepath = os.path.join(subfolder, filename)
# Read the audio signal from the input file
sampling_freq, signal = wavfile.read(filepath)
# Extract the MFCC features
with warnings.catch_warnings():
warnings.simplefilter('ignore')
features_mfcc = mfcc(signal, sampling_freq)
# Append to the variable X
if len(X) == 0:
X = features_mfcc
else:
X = np.append(X, features_mfcc, axis=0)
# Create the HMM model
model = ModelHMM()
# Train the HMM model
model.train(X)
# Save the model for the current word
speech_models.append((model, label))
# Reset the variable
model = None
return speech_models
def predict(self, soundclip):
""" Recognizes the speaker in the sound clip. """
sample_rate, data = wavfile.read(os.path.abspath(soundclip))
ceps = mfcc(data, sample_rate)
fvec = self._mfcc_to_fvec(ceps)
speaker_id = self.recognizer.predict(fvec)[0]
return self.spkr_iton[speaker_id]
