def calculate_acoustic_features(args, waveform):
n_fft = int(args.window*SAMPLE_RATE/1000.0)
hop_length = int(args.step * SAMPLE_RATE / 1000.0)
if 'mfe' == args.feature_type:
if args.backend=='speechpy':
log_cut = 1e-8
spec, energy = mfe(waveform, SAMPLE_RATE, frame_length=args.window*1e-3,
frame_stride=args.step*1e-3, num_filters=args.n_mels, fft_length=n_fft)
if args.energy:
acoustic_features = np.hstack((spec, energy[:, np.newaxis]))
acoustic_features = np.log(acoustic_features + log_cut)
else:
spec = librosa.feature.melspectrogram(y=waveform, sr=SAMPLE_RATE, n_fft=n_fft,
hop_length=hop_length, n_mels=args.n_mels)
acoustic_features = librosa.core.amplitude_to_db(spec).transpose()
if args.energy:
energy = librosa.feature.rmse(y=waveform, frame_length=n_fft, hop_length=hop_length).transpose()
acoustic_features = np.hstack((acoustic_features, energy))
elif 'mfcc' == args.feature_type:
if args.backend=='speechpy':
acoustic_features = mfcc(waveform, SAMPLE_RATE, frame_length=args.window*1e-3,
frame_stride=args.step*1e-3, num_filters=args.n_mels, fft_length=n_fft,
num_cepstral = args.n_mfcc)
else:
acoustic_features = librosa.feature.mfcc(y=waveform, sr=SAMPLE_RATE, n_mfcc=args.n_mfcc,
n_fft=n_fft, hop_length=hop_length, n_mels=args.n_mels).transpose()
if args.energy:
energy = librosa.feature.rmse(y=waveform, frame_length=n_fft, hop_length=hop_length).transpose()
acoustic_features = np.hstack((acoustic_features, energy))
elif 'lyon' == args.feature_type:
waveform /= np.abs(waveform).max()
acoustic_features = lyon_calc.lyon_passive_ear(waveform[:, np.newaxis].astype(np.double),
SAMPLE_RATE, hop_length)
max_val = acoustic_features.max()
if max_val > 0:
acoustic_features /= max_val
acoustic_features = acoustic_features.astype(np.float32)
if args.energy:
energy = librosa.feature.rmse(y=waveform, frame_length=hop_length, hop_length=hop_length).transpose()
energy /= energy.max()
len_delta = acoustic_features.shape[0] - energy.shape[0]
if len_delta > 0:
energy = np.pad(energy, [(0, len_delta), (0, 0)], 'edge')
else:
energy = energy[:acoustic_features.shape[0], :]
acoustic_features = np.hstack((acoustic_features, energy))
else:
raise ValueError('Unexpected features type.')
if args.deltas:
orig_shape = acoustic_features.shape
if args.backend=='speechpy':
acoustic_features = extract_derivative_feature(acoustic_features)
else:
delta = librosa.feature.delta(acoustic_features, axis=0)
ddelta = librosa.feature.delta(acoustic_features, order=2, axis=0)
acoustic_features = np.stack((acoustic_features[:, :, np.newaxis],
delta[:, :, np.newaxis], ddelta[:, :, np.newaxis]), axis=-1)
acoustic_features = np.reshape(acoustic_features, (-1, orig_shape[-1] * 3))
return acoustic_features
def compute_fbank(file, debug=True):
sr, signal = wav.read(file)
if debug:
print('signal shape: ', signal.shape)
# Pre-emphasizing.
signal_preemphasized = processing.preemphasis(signal,
cof=data_config.preemphasis)
# Stacking frames
frames = processing.stack_frames(signal_preemphasized,
sampling_frequency=sr,
frame_length=data_config.window_size,
frame_stride=data_config.hop_size,
zero_padding=True)
# Extracting power spectrum
power_spectrum = processing.power_spectrum(
frames, fft_points=512) # num_frames x fft_length
if debug:
print('power spectrum shape=', power_spectrum.shape)
############# Extract fbanks features #############
log_fbank = feature.lmfe(signal_preemphasized,
sampling_frequency=sr,
frame_length=data_config.window_size,
frame_stride=data_config.hop_size,
num_filters=data_config.num_mels,
fft_length=512,
low_frequency=0,
high_frequency=None) # num_frames x num_filters
if data_config.apply_cmvn:
# Cepstral mean variance normalization.
log_fbank_cmvn = processing.cmvn(log_fbank,
variance_normalization=True)
if debug:
print('fbank(mean + variance normalized) feature shape=',
log_fbank_cmvn.shape)
log_fbank = log_fbank_cmvn # num_frames x num_filters
# Extracting derivative features
log_fbank = feature.extract_derivative_feature(log_fbank)
# print('log fbank feature cube shape=', log_fbank_feature_cube.shape) # num_frames x num_filters x 3
# frameSlice and dowmSampling
# concat_mat = concat_frame(log_fbank)
# log_fbank = subsampling(concat_mat)
# log_fbank = build_LFR_features(log_fbank, data_config.LFR_m, data_config.LFR_n)
if debug:
print('concat & subsample shape=', log_fbank.shape)
return log_fbank
def generate_mfec_features(audio_file_name):
final = []
fs, signal = wav.read(audio_file_name)
mfe = speechpy.feature.mfe(signal,
sampling_frequency=fs,
frame_length=0.02,
frame_stride=0.02,
num_filters=40,
fft_length=512,
low_frequency=0,
high_frequency=None)
mfe_1 = mfe[0][:15]
mfe_final = extract_derivative_feature(mfe_1)
for d1 in mfe_final:
temp_array = []
for d2 in d1:
temp = np.array(d2).reshape(1, 3)
temp_array.append(temp)
final.append(temp_array)
return final
############# Extract MFCC features #############
mfcc = feature.mfcc(signal,
sampling_frequency=fs,
frame_length=0.020,
frame_stride=0.01,
num_filters=40,
fft_length=512,
low_frequency=0,
high_frequency=None)
# Cepstral mean variance normalization.
mfcc_cmvn = processing.cmvn(mfcc, variance_normalization=True)
print('mfcc(mean + variance normalized) feature shape=', mfcc_cmvn.shape)
# Extracting derivative features
mfcc_feature_cube = feature.extract_derivative_feature(mfcc)
print('mfcc feature cube shape=', mfcc_feature_cube.shape)
############# Extract logenergy features #############
logenergy = feature.lmfe(signal,
sampling_frequency=fs,
frame_length=0.020,
frame_stride=0.01,
num_filters=40,
fft_length=512,
low_frequency=0,
high_frequency=None)
logenergy_feature_cube = feature.extract_derivative_feature(logenergy)
print('logenergy features=', logenergy.shape)
if count != 0:
elapsed_time = time.time() - start_time
curr_id = subdir[29:38]
# print(curr_id)
count = count + 1
start_time = time.time()
for file in files:
sound = AudioSegment.from_wav(subdir + "/" + file)
sound = sound.set_channels(1)
sound.export("modified.wav", format="wav")
sample_rate, samples = wavfile.read("modified.wav")
features = lmfe(samples, sample_rate, 0.025, 0.01, 40)
features = extract_derivative_feature(features)
timevar = 100
if features.shape[0] >= timevar:
no_cuts = int(features.shape[0] / timevar)
for i in range(no_cuts):
cut = features[i * timevar:(i * timevar) + timevar:, :, :]
# print("cut: ", cut.shape)
with open(filename2, "a") as myfile:
myfile.write(curr_id + "\n")
with open(filename, "a") as myfile:
for data_slice in cut:
np.savetxt(myfile,
data_slice,
delimiter=',',
newline="\n")