hop_length = dataset['hop_length'][()]
sample_rate = dataset['sample_rate'][()]
#The prediction sliding window will be the array that the model is made to predict new frames from.
mags = X[0]
mags = np.append(mags, Y, axis=0)
print("Dataset Y Shape: ", Y.shape)
print("Mags Shape: ", mags.shape)
print('Synthesising Audio')
#Generate some phases for evey new frame of predicted magnitdues using the phase gen.py library.
phases = phase_gen.gen_phases(mags.shape[0], fftFrameSize, hop_length, sample_rate)
print('Generated New Phases', phases.shape)
#convert all of the predicted magnitudes and generated phases back into samples with the correct hop length.
audio = phase_gen.fft2samples(mags, phases, hop_length)
#this is to stop librosa from saving the .wavs as 64-bit floats with no one can read and normalising it here because librosa's write to wav only normalises floats.
maxv = np.iinfo(np.int16).max
audio_wav = (librosa.util.normalize(audio) * maxv).astype(np.int16)
#Create a unique name and directory for the new audiofiles and write it to wav.
if not os.path.exists('datasets/phased'):
os.makedirs('datasets/phased')
audio_name = dataset_name+'_'+'phased'+'_'+str(it_i)+'.wav'
while os.path.exists('datasets/phased/'+audio_name):
it_i += 1
audio_name = dataset_name+'_'+'phased'+'_'+str(it_i)+'.wav'
if training:
model.fit(train_x, train_y, validation_set=((valid_x, valid_y)),
show_metric=True, batch_size=batch_size, n_epoch=training_iters,
snapshot_epoch=False, snapshot_step=1000, run_id=tf_id,
callbacks=callback)
else:
model.load(cp)
for i in xrange(amount_generated_sequences):
random_index = random.randint(0, (len(train_x) - 1))
impulse = np.array(train_x[random_index])
predicted_magnitudes = impulse
for j in xrange(sequence_max_length):
impulse = np.array(impulse).reshape(1, x_frames.shape[1],
x_frames.shape[2])
prediction = model.predict(impulse)
predicted_magnitudes = np.vstack((predicted_magnitudes, prediction))
impulse = predicted_magnitudes[-sequence_length:]
predicted_magnitudes = np.array(predicted_magnitudes)
print i, predicted_magnitudes.shape
phases = phase_gen.gen_phases(predicted_magnitudes.shape[0], fft_size,
hop_size, sample_rate)
audio = phase_gen.fft2samples(predicted_magnitudes, phases, hop_size)
maxv = np.iinfo(np.int16).max
audio_wav = (librosa.util.normalize(audio) * maxv).astype(np.int16)
audio_name = tf_id + u'_' + unicode(i) + u'.wav'
librosa.output.write_wav(audio_path + u'/' + audio_name, audio_wav,
sample_rate, norm=False)
#For every frame of magnitudes
for i in range(magnitude_t.shape[0]):
#Append the frame to the fft bank.
fft_bank.append(magnitude_t[i])
#Convert the fft bank list into a numpy array.
fft_bank = np.array(fft_bank)
print('Synthesising Audio')
#Generate some phases for evey new frame of predicted magnitdues using the phase gen.py library.
phases = phase_gen.gen_phases(fft_bank.shape[0], fftFrameSize, hop_length,
sample_rate)
print('Generated New Phases', phases.shape)
#convert all of the predicted magnitudes and generated phases back into samples with the correct hop length.
audio = phase_gen.fft2samples(fft_bank, phases, hop_length)
#this is to stop librosa from saving the .wavs as 64-bit floats with no one can read and normalising it here because librosa's write to wav only normalises floats.
maxv = np.iinfo(np.int16).max
audio_wav = (librosa.util.normalize(audio) * maxv).astype(np.int16)
#Add one on to the user defined sequence length.
sequence_length = sequence_length + 1
sequences = []
#Make a bunch of sequences of the FFT bank that are one more than the sequence length.
for i in range(len(fft_bank) - sequence_length):
sequences.append(fft_bank[i:i + sequence_length])
sequences = np.array(sequences)