def process_directory(dir, n_rate):
signal = []
for j, audioname in enumerate(f[dir]):
holder_signal, sr = Audio.read(f'{origin_path}/{dir}/{audioname}',
sr=n_rate)
signal.extend(Audio.trim(holder_signal, 20))
signal = array(signal)
Audio.write(f'{dest_path}/{n_rate}/{dir}.wav', signal, n_rate)
def process_directory(dir, index, library):
signal, rate = Audio.read(f'{path}/{dir}',
sr=sampling_rate,
normalize=True)
signal = np.array(signal)
segment_time = 5
# arredonda o sinal de audio para multiplo de 5
signal = signal[:len(signal) - len(signal) % (rate * segment_time)]
# avalia quantos segmentos têm em uma audio
segments = len(signal) // (rate * segment_time)
augsize = int(augment[0]) if len(augment) > 0 else 0
m = {
'attrs': [],
'labels': [index] * (n_segments or segments) * (1 + augsize),
'classes': [{
index: (n_segments or segments) * (1 + augsize)
}]
}
for i in range(segments):
if n_segments and i >= n_segments:
continue
samples = [Audio.segment(signal, rate, seconds=segment_time, window=i)]
n_mfcc = 13
n_mels = 26
n_fft = 2048
# Janela e overlapping (em amostras)
hop_length = 512
win_length = 1024
# Janela e overlapping (em tempo)
win_len = win_length / rate
win_hop = hop_length / rate
lifter = 22
fmin = 0
fmax = rate / 2
coef_pre_enfase = 0.97
append_energy = 0
if augment:
for _ in range(int(augment[0])):
flag = False
aug = samples[0]
if random.uniform() > 0.5 and 'cut' in augment:
aug = _cut(aug, rate)
flag = True
if random.uniform() > 0.5 and 'noise' in augment:
aug = _noise(aug, rate)
flag = True
if not flag and len(augment) == 3:
if random.uniform() > 0.5:
aug = _cut(aug, rate)
else:
aug = _noise(aug, rate)
samples.append(aug)
for sample_index, sample in enumerate(samples):
if library == 'stft':
attr = np.abs(
np.array(stft(sample, n_fft=n_fft, hop_length=hop_length)))
if library == 'melbanks':
sample = sample[newaxis, :]
melfbanks = MelFilterbanks(sample_rate=rate)
attr = melfbanks(sample)
attr = np.array(attr).T
if library == 'psf':
attr = mfcc(signal=sample,
samplerate=rate,
winlen=win_len,
winstep=win_hop,
numcep=n_mfcc,
nfilt=n_mels,
nfft=n_fft,
lowfreq=fmin,
highfreq=fmax,
preemph=coef_pre_enfase,
ceplifter=lifter,
appendEnergy=append_energy,
winfunc=hann)
attr = np.array(attr)
# Visualization.plot_cepstrals(
# attr, fig_name=f'teste.png')
# Visualization.plot_audio(
# sample, rate, fig_name='./teste.png')
# Audio.write(
# f'portuguese/processed/psf/{dir}_{i}_{sample_index}.wav', sample, rate)
m['attrs'].append(attr.tolist())
del attr
del signal
return m
warped_masked_spectrogram = spec_augment_tensorflow.spec_augment(mfcc)
warped_masked_spectrogram = warped_masked_spectrogram.numpy()
# %%
plt.figure(figsize=(10, 4))
librosa.display.specshow(librosa.power_to_db(warped_masked_spectrogram[0, :, :,
0],
ref=np.max),
y_axis='mel',
fmax=8000,
x_axis='time')
plt.tight_layout()
plt.title('SpecAugmented')
plt.show()
plt.close()
audio_signal = librosa.core.spectrum.griffinlim(
warped_masked_spectrogram[0, :, :, 0])
Audio.write('test/warped_audio.wav', audio_signal, sr)
# %%
plt.figure(figsize=(10, 4))
librosa.display.specshow(librosa.power_to_db(mfcc[0, :, :, 0], ref=np.max),
y_axis='mel',
fmax=8000,
x_axis='time')
plt.tight_layout()
plt.title('MFCC')
plt.show()
plt.close()
audio_signal = librosa.core.spectrum.griffinlim(mfcc[0, :, :, 0])
Audio.write('test/mfcc_audio.wav', audio_signal, sr)
model = load(open(filename_holder + 'model.h5', 'rb'))
signal, rate = librosa.load(args['inferencia'], sr=sampling_rate)
# signal = Audio.trim(signal)
segment_time = 5
signal = signal[:len(signal) - len(signal) % (rate * segment_time)]
segments = len(signal) // (rate * segment_time)
mfcc_audios = []
for i in range(segments):
sample = Audio.segment(signal, rate, seconds=segment_time, window=i)
n_mfcc = 13
n_mels = 26
n_fft = 2048
# Janela e overlapping (em amostras)
hop_length = 512
win_length = 1024
# Janela e overlapping (em tempo)
win_len = win_length / rate
win_hop = hop_length / rate
lifter = 22
fmin = 0
fmax = rate / 2
coef_pre_enfase = 0.97
append_energy = 0
from matplotlib.pyplot import show
from python_speech_features import mfcc
from scipy.signal.windows import hann
from deep_audio import Audio, Visualization
signal, rate = Audio.read(
'inferencia/hugo/Frase 1-1.wav', 24000)
rate = 24000
n_mfcc = 40
n_mels = 40
n_fft = 2048
# Janela e overlapping (em amostras)
hop_length = 512
win_length = 1024
# Janela e overlapping (em tempo)
win_len = win_length / rate
win_hop = hop_length / rate
lifter = 22
fmin = 0
fmax = rate / 2
coef_pre_enfase = 0.97
append_energy = 0
attr = mfcc(
signal=signal,
samplerate=rate,
winlen=win_len,
winstep=win_hop,