def parser(wavefile,num_mfcc,env_noise=None):
try:
y, sr = librosa.load(wavefile, res_type= 'kaiser_fast')
y = prep_data.normalize(y)
#remove the silence at beginning and end of recording
y = prep_data.get_speech_samples(y,sr)
rand_scale = 0.0
#randomly assigns speaker data to 1 (train) 2 (validation) or 3 (test)
if env_noise is not None:
#at random apply varying amounts of environment noise
rand_scale = random.choice([0.0,0.25,0.5,0.75,1.0,1.25])
logging.info("Scale of noise applied: {}".format(rand_scale))
if rand_scale:
#apply *known* environemt noise to signal
total_length = len(y)/sr
envnoise_normalized = prep_data.normalize(env_noise)
envnoise_scaled = prep_data.scale_noise(envnoise_normalized,rand_scale)
envnoise_matched = prep_data.match_length(envnoise_scaled,sr,total_length)
if len(envnoise_matched) != len(y):
diff = int(len(y) - len(envnoise_matched))
if diff < 0:
envnoise_matched = envnoise_matched[:diff]
else:
envnoise_matched = np.append(envnoise_matched,np.zeros(diff,))
y += envnoise_matched
mfccs = librosa.feature.mfcc(y, sr, n_mfcc=num_mfcc,hop_length=int(0.010*sr),n_fft=int(0.025*sr))
return mfccs, sr, rand_scale
except EOFError as error:
logging.exception('def parser() resulted in {} for the file: {}'.format(error,wavefile))
except ValueError as ve:
logging.exception("Error occured ({}) with the file {}".format(ve,wavefile))
return None, None, None
def apply_noise(y, sr, wavefile):
#at random apply varying amounts of environment noise
rand_scale = random.choice([0.0, 0.25, 0.5, 0.75])
if rand_scale > 0.0:
#apply *known* environemt noise to signal
total_length = len(y) / sr
y_noise, sr = librosa.load(wavefile, sr=16000)
envnoise_normalized = prep_data_vad_noise.normalize(y_noise)
envnoise_scaled = prep_data_vad_noise.scale_noise(
envnoise_normalized, rand_scale)
envnoise_matched = prep_data_vad_noise.match_length(
envnoise_scaled, sr, total_length)
if len(envnoise_matched) != len(y):
diff = int(len(y) - len(envnoise_matched))
if diff < 0:
envnoise_matched = envnoise_matched[:diff]
else:
envnoise_matched = np.append(envnoise_matched,
np.zeros(diff, ))
y += envnoise_matched
return y