def generate_audio(): # compute receptive field width learnable_steps = 1 batch_size = 1 num_layers = len(params.residual_conv_channels) receptive_steps_per_unit = params.residual_conv_filter_width ** num_layers receptive_steps = (receptive_steps_per_unit - 1) * params.residual_num_blocks + 1 target_width = learnable_steps input_width = receptive_steps # to compute all learnable targets input_width += learnable_steps - 1 ## padding for causal conv block input_width += len(params.causal_conv_channels) # quantized signals generated by WaveNet generated_quantized_audio = np.zeros((input_width, ), dtype=np.int32) start = time.time() for time_step in xrange(9): # quantized signals in receptive field padded_quantized_x_batch = generated_quantized_audio[-input_width:].reshape((1, -1)) # convert to image padded_x_batch = data.onehot_pixel_image(padded_quantized_x_batch, quantization_steps=params.quantization_steps) # generate next signal softmax = wavenet.forward_one_step(padded_x_batch, softmax=True, return_numpy=True) softmax = softmax[0, :, 0, -1] generated_quantized_signal = np.argmax(softmax) generated_quantized_audio = np.append(generated_quantized_audio, [generated_quantized_signal], axis=0) print generated_quantized_signal, print generated_quantized_audio print time.time() - start wavenet.prev_causal_outputs = None wavenet.prev_residual_outputs_out = None wavenet.prev_residual_outputs_z = None generated_quantized_audio = np.zeros((input_width, ), dtype=np.int32) start = time.time() for time_step in xrange(9): # quantized signals in receptive field padded_quantized_x_batch = generated_quantized_audio[-input_width:].reshape((1, -1)) # convert to image padded_x_batch = data.onehot_pixel_image(padded_quantized_x_batch, quantization_steps=params.quantization_steps) # generate next signal softmax = wavenet._forward_one_step(padded_x_batch, softmax=True, return_numpy=True) softmax = softmax[0, :, 0, -1] generated_quantized_signal = np.argmax(softmax) generated_quantized_audio = np.append(generated_quantized_audio, [generated_quantized_signal], axis=0) print generated_quantized_signal, print generated_quantized_audio print time.time() - start
def generate_audio(sampling_rate=48000, generate_sec=1, remove_silence_frames=False):
batch_size = 1
# compute required input width
num_layers = len(params.residual_conv_channels)
receptive_steps_per_unit = params.residual_conv_filter_width ** num_layers
receptive_steps = (receptive_steps_per_unit - 1) * params.residual_num_blocks + 1
input_width = receptive_steps
# padding for causal conv block
input_width += len(params.causal_conv_channels)
# quantized signals generated by WaveNet
generated_quantized_audio = np.zeros((input_width, ), dtype=np.int32)
start_time = time.time()
for time_step in xrange(1, int(sampling_rate * generate_sec)):
# quantized signals in receptive field
padded_quantized_x_batch = generated_quantized_audio[-input_width:].reshape((1, -1))
# convert to image
padded_x_batch = data.onehot_pixel_image(padded_quantized_x_batch, quantization_steps=params.quantization_steps)
# generate next signal
if args.use_faster_wavenet:
softmax = wavenet._forward_one_step(padded_x_batch, softmax=True, return_numpy=True)
else:
softmax = wavenet.forward_one_step(padded_x_batch, softmax=True, return_numpy=True)
softmax = softmax[0, :, 0, -1]
generated_quantized_signal = np.random.choice(np.arange(params.quantization_steps), p=softmax)
if generated_quantized_signal == 0 and remove_silence_frames:
pass
else:
generated_quantized_audio = np.append(generated_quantized_audio, [generated_quantized_signal], axis=0)
if time_step % 10 == 0:
sys.stdout.write("\rgenerating {:.2f} msec / {:.2f} msec".format(time_step * 1000.0 / sampling_rate, generate_sec * 1000.0))
sys.stdout.flush()
print "\ndone in {:.3f} sec".format(time.time() - start_time)
# remove zero paddings
generated_quantized_audio = generated_quantized_audio[input_width:]
try:
os.mkdir(args.generate_dir)
except:
pass
filename = "{}/generated.wav".format(args.generate_dir)
data.save_audio_file(filename, generated_quantized_audio, params.quantization_steps, format="16bit_pcm", sampling_rate=sampling_rate)
def generate_audio(sampling_rate=48000, generate_sec=1, remove_silence_frames=False):
# compute required input width
num_layers = len(params.residual_conv_channels)
receptive_steps_per_unit = params.residual_conv_filter_width ** num_layers
receptive_steps = (receptive_steps_per_unit - 1) * params.residual_num_blocks + 1
input_width = receptive_steps
# add paddings of causal conv block
input_width += len(params.causal_conv_channels)
# pad with silence signals
generated_signals = np.full((input_width, ), 127, dtype=np.int32)
start_time = time.time()
for time_step in xrange(1, int(sampling_rate * generate_sec)):
# signals in receptive field
input_signals = generated_signals[-input_width:].reshape((1, -1))
# convert to image
input_signals = data.onehot_pixel_image(input_signals, quantization_steps=params.quantization_steps)
# generate next signal
if args.fast:
softmax = wavenet._forward_one_step(input_signals, apply_softmax=True, as_numpy=True)
else:
softmax = wavenet.forward_one_step(input_signals, apply_softmax=True, as_numpy=True)
softmax = softmax[0, :, 0, -1]
signal = np.random.choice(np.arange(params.quantization_steps), p=softmax)
if signal == 127 and remove_silence_frames:
pass
else:
generated_signals = np.append(generated_signals, [signal], axis=0)
if time_step % 10 == 0:
sys.stdout.write("\rgenerating {:.2f} msec / {:.2f} msec".format(time_step * 1000.0 / sampling_rate, generate_sec * 1000.0))
sys.stdout.flush()
print "\ndone in {:.3f} sec".format(time.time() - start_time)
# remove paddings
generated_signals = generated_signals[input_width:]
try:
os.mkdir(args.output_dir)
except:
pass
filename = "{}/generated.wav".format(args.output_dir)
data.save_audio_file(filename, generated_signals, params.quantization_steps, format="16bit_pcm", sampling_rate=sampling_rate)
def generate_audio(): receptive_field_width_steps = 5 batch_size = 1 max_dilation = max(params.residual_conv_dilations) target_width = receptive_field_width_steps padded_input_width = receptive_field_width_steps + max_dilation # quantized signals generated by WaveNet generated_quantized_audio = np.mod(np.arange(1, padded_input_width + 1), 6).astype(np.int32) start = time.time() for time_step in xrange(500): # quantized signals in receptive field padded_quantized_x_batch = generated_quantized_audio[-padded_input_width:].reshape((1, -1)) # convert to image padded_x_batch = data.onehot_pixel_image(padded_quantized_x_batch, quantized_channels=params.audio_channels) # generate next signal softmax = wavenet.forward_one_step(padded_x_batch, softmax=True, return_numpy=True) softmax = softmax[0, :, 0, -1] generated_quantized_signal = np.argmax(softmax) generated_quantized_audio = np.append(generated_quantized_audio, [generated_quantized_signal], axis=0) print generated_quantized_audio print time.time() - start wavenet.prev_causal_outputs = None wavenet.prev_residual_outputs_out = None wavenet.prev_residual_outputs_z = None generated_quantized_audio = np.mod(np.arange(1, padded_input_width + 1), 6).astype(np.int32) start = time.time() for time_step in xrange(500): # quantized signals in receptive field padded_quantized_x_batch = generated_quantized_audio[-padded_input_width:].reshape((1, -1)) # convert to image padded_x_batch = data.onehot_pixel_image(padded_quantized_x_batch, quantized_channels=params.audio_channels) # generate next signal softmax = wavenet._forward_one_step(padded_x_batch, softmax=True, return_numpy=True) softmax = softmax[0, :, 0, -1] generated_quantized_signal = np.argmax(softmax) generated_quantized_audio = np.append(generated_quantized_audio, [generated_quantized_signal], axis=0) print generated_quantized_audio print time.time() - start
def generate_audio(receptive_field_width_ms=25, sampling_rate=48000, generate_duration_sec=1):
# e.g.
# 48000 Hz * 0.25 = 12000 time steps (= 250 milliseconds receptive field)
receptive_steps = int(sampling_rate * receptive_field_width_ms / 1000.0)
# compute required input width
batch_size = 1
max_dilation = max(params.residual_conv_dilations)
target_width = receptive_steps
padded_input_width = receptive_steps + max_dilation * (params.residual_conv_kernel_width - 1)
# quantized signals generated by WaveNet
generated_quantized_audio = np.zeros((padded_input_width, ), dtype=np.int32)
start_time = time.time()
for time_step in xrange(1, int(sampling_rate * generate_duration_sec)):
# quantized signals in receptive field
padded_quantized_x_batch = generated_quantized_audio[-padded_input_width:].reshape((1, -1))
# convert to image
padded_x_batch = data.onehot_pixel_image(padded_quantized_x_batch, quantized_channels=params.audio_channels)
# generate next signal
if args.use_faster_wavenet:
softmax = wavenet._forward_one_step(padded_x_batch, softmax=True, return_numpy=True)
else:
softmax = wavenet.forward_one_step(padded_x_batch, softmax=True, return_numpy=True)
softmax = softmax[0, :, 0, -1]
generated_quantized_signal = np.random.choice(np.arange(params.audio_channels), p=softmax)
generated_quantized_audio = np.append(generated_quantized_audio, [generated_quantized_signal], axis=0)
if time_step % 10 == 0:
sys.stdout.write("\rgenerating {:.2f} msec / {:.2f} msec".format(time_step * 1000.0 / sampling_rate, generate_duration_sec * 1000.0))
sys.stdout.flush()
print "\ndone in {:.3f} sec".format(time.time() - start_time)
# remove zero paddings
generated_quantized_audio = generated_quantized_audio[padded_input_width:]
try:
os.mkdir(args.generate_dir)
except:
pass
filename = "{}/generated.wav".format(args.generate_dir)
data.save_audio_file(filename, generated_quantized_audio, params.audio_channels, format="16bit_pcm", sampling_rate=sampling_rate)