def test_ubm_var_channel():
ubm = GMM.load('model/ubm.mixture-32.person-20.immature.model')
train_duration = 8.
nr_test = 5
test_duration = 3.
audio_files = ['xinyu.vad.wav', 'wyx.wav']
X_train, y_train, X_test, y_test = [], [], [], []
for audio_file in audio_files:
fs, signal = wavfile.read(audio_file)
signal = monotize_signal(signal)
train_len = int(fs * train_duration)
test_len = int(fs * test_duration)
X_train.append(mix_feature((fs, signal[:train_len])))
y_train.append(audio_file)
for i in range(nr_test):
start = random.randint(train_len, len(signal) - test_len)
X_test.append(mix_feature((fs, signal[start:start + train_len])))
y_test.append(audio_file)
gmmset = GMMSet(32, ubm=ubm)
gmmset.fit(X_train, y_train)
y_pred = gmmset.predict_with_reject(X_test)
for i in range(len(y_pred)):
print(y_test[i], y_pred[i], '' if y_test[i] == y_pred[i] else 'wrong')
for imposter_audio_file in map(lambda x: 'test-{}.wav'.format(x),
range(5)):
fs, signal = wavfile.read(imposter_audio_file)
signal = monotize_signal(signal)
imposter_x = mix_feature((fs, signal))
print(gmmset.predict_one_with_rejection(imposter_x))
def load_gmmset(labels, nr_person):
gmmset = GMMSet(concurrency=8)
for fpath in sorted(glob.glob('model.new-mfcc/*')):
fname = os.path.basename(fpath)
base = fname[:fname.find('.')]
if base not in labels:
continue
if fname.endswith("32.model"):
print(base, fname)
gmmset.load_gmm(base, fpath)
return gmmset
def load_gmmset(labels, nr_person):
gmmset = GMMSet(concurrency=8)
for fpath in sorted(glob.glob('model.new-mfcc/*')):
fname = os.path.basename(fpath)
base = fname[:fname.find('.')]
if base not in labels:
continue
if fname.endswith("32.model"):
print base, fname
gmmset.load_gmm(base, fpath)
return gmmset
def test_ubm_var_channel():
ubm = GMM.load('model/ubm.mixture-32.person-20.immature.model')
train_duration = 8.
nr_test = 5
test_duration = 3.
audio_files = ['xinyu.vad.wav', 'wyx.wav']
X_train, y_train, X_test, y_test = [], [], [], []
for audio_file in audio_files:
fs, signal = wavfile.read(audio_file)
signal = monotize_signal(signal)
train_len = int(fs * train_duration)
test_len = int(fs * test_duration)
X_train.append(mix_feature((fs, signal[:train_len])))
y_train.append(audio_file)
for i in range(nr_test):
start = random.randint(train_len, len(signal) - test_len)
X_test.append(mix_feature((fs, signal[start:start+train_len])))
y_test.append(audio_file)
gmmset = GMMSet(32, ubm=ubm)
gmmset.fit(X_train, y_train)
y_pred = gmmset.predict_with_reject(X_test)
for i in xrange(len(y_pred)):
print y_test[i], y_pred[i], '' if y_test[i] == y_pred[i] else 'wrong'
for imposter_audio_file in map(
lambda x: 'test-{}.wav'.format(x), range(5)):
fs, signal = wavfile.read(imposter_audio_file)
signal = monotize_signal(signal)
imposter_x = mix_feature((fs, signal))
print gmmset.predict_one_with_rejection(imposter_x)
def __init__(self):
self.features = defaultdict(list)
self.gmmset = GMMSet()
self.vad = VAD()
class ModelInterface(object):
UBM_MODEL_FILE = 'model/ubm.mixture-32.utt-300.model'
def __init__(self):
self.features = defaultdict(list)
self.gmmset = GMMSet()
self.vad = VAD()
def init_noise(self, fs, signal):
"""
init vad from environment noise
"""
self.vad.init_noise(fs, signal)
def filter(self, fs, signal):
"""
use VAD (voice activity detection) to filter out silence part of a signal
"""
ret, intervals = self.vad.filter(fs, signal)
orig_len = len(signal)
if len(ret) > orig_len / 3:
# signal is filtered by VAD
return ret
return np.array([])
def enroll(self, name, fs, signal):
"""
add the signal to this person's training dataset
name: person's name
"""
feat = mix_feature((fs, signal))
self.features[name].extend(feat)
def _get_gmm_set(self):
if os.path.isfile(self.UBM_MODEL_FILE):
try:
from gmmset import GMMSetPyGMM
if GMMSet is GMMSetPyGMM:
return GMMSet(ubm=GMM.load(self.UBM_MODEL_FILE))
except Exception as e:
print "Warning: failed to import gmmset. You may forget to compile gmm:"
print e
print "Try running `make -C src/gmm` to compile gmm module."
print "But gmm from sklearn will work as well! Using it now!"
return GMMSet()
return GMMSet()
def train(self):
self.gmmset = self._get_gmm_set()
start = time.time()
print "Start training..."
for name, feats in self.features.iteritems():
self.gmmset.fit_new(feats, name)
print time.time() - start, " seconds"
def predict(self, fs, signal):
"""
return a label (name)
"""
try:
feat = mix_feature((fs, signal))
except Exception as e:
print tb.format_exc()
return None
return self.gmmset.predict_one(feat)
def predict_with_score(self, fs, signal):
"""
return a label (name)
"""
try:
feat = mix_feature((fs, signal))
except Exception as e:
print tb.format_exc()
return None
# gmmset = GMMSet() = gmmset.GMMSetPyGMM
return self.gmmset.predict_one_with_score(feat)
def dump(self, fname):
""" dump all models to file"""
self.gmmset.before_pickle()
with open(fname, 'wb') as f:
pickle.dump(self, f, -1)
self.gmmset.after_pickle()
@staticmethod
def load(fname):
""" load from a dumped model file"""
with open(fname, 'rb') as f:
R = pickle.load(f)
R.gmmset.after_pickle()
return R
def __init__(self):
self.features = defaultdict(list)
self.gmmset = GMMSet()
self.vad = VAD()
class ModelInterface(object):
UBM_MODEL_FILE = 'model/ubm.mixture-32.utt-300.model'
def __init__(self):
self.features = defaultdict(list)
self.gmmset = GMMSet()
self.vad = VAD()
def init_noise(self, fs, signal):
"""
init vad from environment noise
"""
self.vad.init_noise(fs, signal)
def filter(self, fs, signal):
"""
use VAD (voice activity detection) to filter out silence part of a signal
"""
ret, intervals = self.vad.filter(fs, signal)
orig_len = len(signal)
if len(ret) > orig_len / 3:
# signal is filtered by VAD
return ret
return np.array([])
def enroll(self, name, fs, signal):
"""
add the signal to this person's training dataset
name: person's name
"""
feat = mix_feature((fs, signal))
self.features[name].extend(feat)
def _get_gmm_set(self):
if os.path.isfile(self.UBM_MODEL_FILE):
try:
from gmmset import GMMSetPyGMM
if GMMSet is GMMSetPyGMM:
return GMMSet(ubm=GMM.load(self.UBM_MODEL_FILE))
except Exception as e:
print "Warning: failed to import gmmset. You may forget to compile gmm:"
print e
print "Try running `make -C src/gmm` to compile gmm module."
print "But gmm from sklearn will work as well! Using it now!"
return GMMSet()
return GMMSet()
def train(self):
self.gmmset = self._get_gmm_set()
start = time.time()
print "Start training..."
for name, feats in self.features.iteritems():
self.gmmset.fit_new(feats, name)
print time.time() - start, " seconds"
def predict(self, fs, signal):
"""
return a label (name)
"""
try:
feat = mix_feature((fs, signal))
except Exception as e:
print tb.format_exc()
return None
return self.gmmset.predict_one(feat)
def dump(self, fname):
""" dump all models to file"""
self.gmmset.before_pickle()
with open(fname, 'w') as f:
pickle.dump(self, f, -1)
self.gmmset.after_pickle()
@staticmethod
def load(fname):
""" load from a dumped model file"""
with open(fname, 'r') as f:
R = pickle.load(f)
R.gmmset.after_pickle()
return R
