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 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 get_mfcc(self, audio_path):
(sr, sig) = wav.read(audio_path)
if len(sig.shape) > 1:
sig = sig[:, 0]
cleansig = remove_silence(sr, sig)
mfcc_vecs = mfcc(cleansig, sr, numcep=19)
mfcc_delta = librosa.feature.delta(mfcc_vecs.T)
mfcc_delta2 = librosa.feature.delta(mfcc_vecs.T, order=2)
feats = np.vstack([mfcc_vecs.T, mfcc_delta, mfcc_delta2])
return feats.T
def enroll(self, name, signal, fs = 44100):
signal_new = remove_silence(fs, signal)
hop_length = np.min([0.016 * fs, 512])
mfcc = librosa.feature.mfcc(y = signal_new, sr = fs, n_mfcc = 15, hop_length = hop_length)
mfcc = mfcc.T
mu = np.mean(mfcc, axis = 0)
sigma = np.std(mfcc, axis = 0)
feature = (mfcc - mu) / sigma
self.features.append(feature)
self.classes.append(name)
def predict(self, signal, fs = 44100):
signal_new = remove_silence(fs, signal)
# if len(signal_new) < len(signal) / 4:
# return "Silence"
hop_length = np.min([0.016 * fs, 512])
mfcc = librosa.feature.mfcc(y = signal_new, sr = fs, n_mfcc = 15, hop_length = hop_length)
mfcc = mfcc.T
mu = np.mean(mfcc, axis = 0)
sigma = np.std(mfcc, axis = 0)
feature = (mfcc - mu) / sigma
return self.gmmset.predict_one(feature)
def get_mfcc(file_path, noise):
fs, y = wavfile.read(to_mono(file_path)) # read .wav file
y = remove_silence(fs, y)
if not np.issubdtype(y.dtype, np.floating):
y = [np.float32(i) for i in y]
y = np.array(y)
if len(noise) > 0:
if not np.issubdtype(noise.dtype, np.floating):
noise = [np.float32(i) for i in noise]
noise = np.array(noise)
y = nr.reduce_noise(audio_clip=np.array(y), noise_clip=np.array(noise), verbose=False)
mfcc = extract(fs, y)
return mfcc
wavfile.write(NOISE_WAV, fs, signal)
os.system("sox {0} -n noiseprof {1}".format(NOISE_WAV, NOISE_MODEL))
def filter(self, fs, signal):
rand = r.randint(1, 100000)
fname = "/tmp/tmp{0}.wav".format(rand)
signal = monophonic(signal)
wavfile.write(fname, fs, signal)
fname_clean = "/tmp/tmp{0}-clean.wav".format(rand)
os.system("sox {0} {1} noisered {2} {3}".format(
fname, fname_clean, NOISE_MODEL, THRES))
fs, signal = wavfile.read(fname_clean)
signal = monophonic(signal)
os.remove(fname)
os.remove(fname_clean)
return signal
if __name__ == "__main__":
fs, bg = wavfile.read(sys.argv[1])
nr = NoiseReduction()
nr.init_noise(fs, bg)
fs, sig = wavfile.read(sys.argv[2])
vaded = nr.filter(fs, sig)
wavfile.write('vaded.wav', fs, vaded)
removed = remove_silence(fs, vaded)
wavfile.write("removed.wav", fs, removed)
def enroll(self, name, signal, fs=44100):
if len(signal.shape) > 1:
signal = signal[:, 0]
signal_new = remove_silence(fs, signal)
mfcc_vecs = mfcc(signal_new, fs, numcep=15)
self.enroll_feat(name, mfcc_vecs)
def enroll(self, name, signal, fs = 44100):
if len(signal.shape) > 1:
signal = signal[:, 0]
signal_new = remove_silence(fs, signal)
mfcc_vecs = mfcc(signal_new, fs, numcep = 15)
self.enroll_feat(name, mfcc_vecs)