def test_norm_amp(base_filenames):
for f in base_filenames:
print(f)
if f.startswith('silence'):
continue
wavpath = f + '.wav'
mfcc = Mfcc(wavpath,
min_freq=0,
max_freq=8000,
num_coeffs=1,
win_len=0.025,
time_step=0.01,
num_filters=20,
use_power=True)
mfcc.norm_amp([(0, 1)])
def get_features(self, path):
mfcc = Mfcc(path, self.freq_lims, self._num_coeffs, 0.025,
0.01, num_filters = 26, use_power = True, deltas = True)
pitch = Pitch(path, self.time_step, (75,600))
pitch.process()
harmonicity = Harmonicity(path, self.time_step, 75)
harmonicity.process()
return mfcc, pitch, harmonicity
def test(self):
for f in filenames:
print(f)
if f.startswith('silence'):
continue
wavpath = os.path.join(TEST_DIR,f+'.wav')
matpath = os.path.join(TEST_DIR,f+'_mfcc.mat')
if not os.path.exists(matpath):
continue
m = loadmat(matpath)
mfcc = Mfcc(wavpath,self.freq_lims,self.numCC,self.winLen,
self.timeStep,num_filters=self.num_filters,
use_power=True
)
pspec, aspec = mfcc.process(debug=True)
#assert_array_almost_equal(m['pspectrum'].T,pspec,decimal=4)
#assert_array_almost_equal(m['aspectrum'].T,aspec,decimal=4)
assert_array_almost_equal(m['cepstra'].T,mfcc.to_array())
def test(self):
for f in filenames:
print(f)
if f.startswith('silence'):
continue
wavpath = os.path.join(TEST_DIR, f + '.wav')
matpath = os.path.join(TEST_DIR, f + '_mfcc.mat')
if not os.path.exists(matpath):
continue
m = loadmat(matpath)
mfcc = Mfcc(wavpath,
self.freq_lims,
self.numCC,
self.winLen,
self.timeStep,
num_filters=self.num_filters,
use_power=True)
pspec, aspec = mfcc.process(debug=True)
#assert_array_almost_equal(m['pspectrum'].T,pspec,decimal=4)
#assert_array_almost_equal(m['aspectrum'].T,aspec,decimal=4)
assert_array_almost_equal(m['cepstra'].T, mfcc.to_array())
def test(base_filenames):
for f in base_filenames:
print(f)
if f.startswith('silence'):
continue
wavpath = f + '.wav'
matpath = f + '_mfcc.mat'
if not os.path.exists(matpath):
continue
m = loadmat(matpath)
mfcc = Mfcc(wavpath,
min_freq=0,
max_freq=8000,
num_coeffs=13,
win_len=0.025,
time_step=0.01,
num_filters=20,
use_power=True)
mfcc.process()
#assert_array_almost_equal(m['pspectrum'].T,pspec,decimal=4)
#assert_array_almost_equal(m['aspectrum'].T,aspec,decimal=4)
assert_array_almost_equal(m['cepstra'].T, mfcc.to_array())
def test_deltas(base_filenames):
for f in base_filenames:
print(f)
if f.startswith('silence'):
continue
wavpath = f + '.wav'
mfcc = Mfcc(wavpath,
min_freq=0,
max_freq=8000,
num_coeffs=13,
win_len=0.025,
time_step=0.01,
num_filters=20,
use_power=False,
deltas=True)
max_thresh = 0.05
segs = []
for interval in word_tier:
if interval.mark == '':
continue
print(interval.mark, interval.minTime, interval.maxTime)
outpath = os.path.join(temp_wav_dir, interval.mark + '.wav')
extract_audio(wav_path,
outpath,
interval.minTime,
interval.maxTime,
padding=padding)
rep = Mfcc(outpath,
freq_lims=(80, 7800),
num_coeffs=12,
win_len=0.025,
time_step=0.01)
rep.is_windowed = True
duration = interval.maxTime - interval.minTime
thresh = unnorm(norm(duration, min_duration, max_duration), min_thresh,
max_thresh)
rep.segment(threshold=thresh)
print(sorted(rep._segments.keys()))
padded_begin = interval.minTime - padding
if padded_begin < 0:
padded_begin = 0
for k in sorted(rep._segments.keys()):
with open(os.path.join(temp_mfcc_dir, '{}.mfcc'.format(seg_ind)),
'wb') as fh:
pickle.dump(rep[k[0], k[1]], fh)
continue
durations.append(interval.maxTime - interval.minTime)
max_duration = max(durations)
min_duration = min(durations)
min_thresh = 0.01
max_thresh = 0.05
segs = []
for interval in word_tier:
if interval.mark == '':
continue
print(interval.mark, interval.minTime, interval.maxTime)
outpath = os.path.join(temp_wav_dir, interval.mark + '.wav')
extract_audio(wav_path, outpath, interval.minTime, interval.maxTime, padding = padding)
rep = Mfcc(outpath, freq_lims = (80, 7800), num_coeffs = 12, win_len = 0.025, time_step = 0.01)
rep.is_windowed = True
duration = interval.maxTime - interval.minTime
thresh = unnorm(norm(duration, min_duration, max_duration), min_thresh, max_thresh)
rep.segment(threshold = thresh)
print(sorted(rep._segments.keys()))
padded_begin = interval.minTime - padding
if padded_begin < 0:
padded_begin = 0
for k in sorted(rep._segments.keys()):
with open(os.path.join(temp_mfcc_dir, '{}.mfcc'.format(seg_ind)), 'wb') as fh:
pickle.dump(rep[k[0],k[1]], fh)
with open(os.path.join(temp_mean_dir, '{}.mean'.format(seg_ind)), 'wb') as fh:
pickle.dump(rep._segments[k], fh)
segs.append(str(seg_ind))
seg_ind += 1
def test(self, debug = False, dialect = 'timit', norm_amp = True, classifier = 'hmm'):
print('begin testing')
#assume testing on TIMIT
if debug:
with open('debug.txt','w') as f:
pass
words = []
phones = []
wavs = []
if dialect == 'timit':
path = TIMIT_DIR
test_dir = os.path.join(path, 'TEST')
segment_set = TIMIT_SEGMENT_SET
wrdExt = '.wrd'
phnExt = '.phn'
elif dialect == 'buckeye':
test_dir = BUCKEYE_DIR
segment_set = BUCKEYE_SEGMENT_SET
wrdExt = '.words'
phnExt = '.phones'
for root, subdirs, files in os.walk(test_dir):
for f in files:
if f.lower().endswith(wrdExt):
words.append(os.path.join(root,f))
elif f.lower().endswith(phnExt):
phones.append(os.path.join(root,f))
elif f.lower().endswith('.wav'):
wavs.append(os.path.join(root,f))
dialogs = align_dialog_info(words, phones,wavs)
hits = 0
total = 0
cats = sorted(self._priors.keys())
cats2ind = {x:i for i,x in enumerate(cats)}
conf_matrix = np.zeros((len(cats),len(cats)))
for (i,(d,info)) in enumerate(dialogs.items()):
print(i,len(dialogs))
mfcc = Mfcc(info['wav'], (80,7800), self._num_coeffs, 0.025,
0.01, num_filters = 26, use_power = True)
if norm_amp:
mfcc.norm_amp(*self._ranges[0])
phones = read_phones(info['phones'],sr=mfcc.sampling_rate, dialect=dialect)
if debug:
with open('debug.txt','a') as f:
f.write('{}\n{}\n\n'.format(info['wav'],' '.join([x[0] for x in phones])))
for p in phones:
#if p[0] in segment_set['V']:
# phone_class = 'V'
#else:
# phone_class = 'NV'
for k,v in segment_set.items():
if p[0] in v:
phone_class = k
break
else:
continue
if debug:
with open('debug.txt','a') as f:
f.write('{}\n\n'.format(p))
coeffs = mfcc[p[1],p[2]]
for frame in coeffs:
predicted = self.predict(frame)
colInd = cats2ind[predicted]
rowInd = cats2ind[phone_class]
conf_matrix[rowInd,colInd] += 1
if predicted == phone_class:
hits += 1
total += 1
if debug:
with open('debug.txt','a') as f:
f.write('{}\t{}\t{}\n'.format(phone_class,predicted,' '.join(map(str,frame))))
print('Current accuracy:',hits/total)
print('Accuracy:',hits/total)
print(cats)
print(conf_matrix)