def compute_mfcc_feats(input_path, output_path,
compress, compression_method, write_num_frames, **kwargs):
mfcc_args = MFCC.filter_args(**kwargs)
mfcc = MFCC(**mfcc_args)
if mfcc.input_step == 'wave':
input_args = AR.filter_args(**kwargs)
reader = AR(input_path, **input_args)
else:
input_args = DRF.filter_args(**kwargs)
reader = DRF.create(input_path, **input_args)
writer = DWF.create(output_path, scp_sep=' ',
compress=compress,
compression_method=compression_method)
if write_num_frames is not None:
f_num_frames = open(write_num_frames, 'w')
for data in reader:
if mfcc.input_step == 'wave':
key, x, fs = data
else:
key, x = data
logging.info('Extracting MFCC for %s' % (key))
t1 = time.time()
y = mfcc.compute(x)
dt = (time.time() - t1)*1000
rtf = mfcc.frame_shift*y.shape[0]/dt
logging.info('Extracted MFCC for %s num-frames=%d elapsed-time=%.2f ms. real-time-factor=%.2f' %
(key, y.shape[0], dt, rtf))
writer.write([key], [y])
if write_num_frames is not None:
f_num_frames.write('%s %d\n' % (key, y.shape[0]))
mfcc.reset()
if write_num_frames is not None:
f_num_frames.close()
def compute_mfcc_feats(input_path, output_path, compress, compression_method,
write_num_frames, use_gpu, nn_model_path, chunk_size,
context, **kwargs):
#open device
if use_gpu and torch.cuda.is_available():
logging.info('CUDA_VISIBLE_DEVICES=%s' %
os.environ['CUDA_VISIBLE_DEVICES'])
logging.info('init gpu device')
device = torch.device('cuda', 0)
torch.tensor([0]).to(device)
else:
logging.info('init cpu device')
device = torch.device('cpu')
mfcc_args = MFCC.filter_args(**kwargs)
mfcc = MFCC(**mfcc_args)
# PUT YOUR NNET MODEL HERE!!!!
enhancer = CAN(num_channels=45)
enhancer.load_state_dict(
torch.load(nn_model_path, map_location=device)['state_dict'])
enhancer.to(device)
enhancer.eval()
if mfcc.input_step == 'wave':
input_args = AR.filter_args(**kwargs)
reader = AR(input_path, **input_args)
else:
input_args = DRF.filter_args(**kwargs)
reader = DRF.create(input_path, **input_args)
writer = DWF.create(output_path,
scp_sep=' ',
compress=compress,
compression_method=compression_method)
if write_num_frames is not None:
f_num_frames = open(write_num_frames, 'w')
for data in reader:
if mfcc.input_step == 'wave':
key, x, fs = data
else:
key, x = data
logging.info('Extracting filter-banks for %s' % (key))
t1 = time.time()
y = mfcc.compute(x)
#we apply dummy identity network to fb
logging.info('Running enhancement network')
y = apply_nnet(y, enhancer, chunk_size, context, device)
dt = (time.time() - t1) * 1000
rtf = mfcc.frame_shift * y.shape[0] / dt
logging.info(
'Extracted filter-banks for %s num-frames=%d elapsed-time=%.2f ms. real-time-factor=%.2f'
% (key, y.shape[0], dt, rtf))
writer.write([key], [y])
if write_num_frames is not None:
f_num_frames.write('%s %d\n' % (key, y.shape[0]))
mfcc.reset()
if write_num_frames is not None:
f_num_frames.close()
def compute_mfcc_feats(input_path, output_path,
compress, compression_method, write_num_frames,
use_gpu, nn_model_path, chunk_size, context,
**kwargs):
#open device
if use_gpu and torch.cuda.is_available():
os.environ['CUDA_DEVICE_ORDER']='PCI_BUS_ID'
max_tries = 100
for g in range(max_tries):
try:
gpu_ids = find_free_gpus()
os.environ['CUDA_VISIBLE_DEVICES'] = gpu_ids
logging.info('CUDA_VISIBLE_DEVICES=%s' % os.environ['CUDA_VISIBLE_DEVICES'])
logging.info('init gpu device')
device = torch.device('cuda', 0)
torch.tensor([0]).to(device)
break
except:
if g < max_tries-1:
logging.info('failing init gpu, trying again')
time.sleep(10)
else:
logging.info('failing init gpu, using cpu')
device = torch.device('cpu')
else:
logging.info('init cpu device')
device = torch.device('cpu')
mfcc_args1 = MFCC.filter_args(**kwargs)
mfcc_args2 = copy.deepcopy(mfcc_args1)
mfcc_args1['output_step'] = 'logfb'
mfcc_args2['input_step'] = 'logfb'
print(kwargs)
print(mfcc_args1)
print(mfcc_args2)
mfcc1 = MFCC(**mfcc_args1)
mfcc2 = MFCC(**mfcc_args2)
mvn = MVN(norm_var=False, left_context=150, right_context=150)
# PUT YOUR NNET MODEL HERE!!!!
enhancer = CGN()
#enhancer.load_state_dict(torch.load(nn_model_path, map_location=device)['state_dict'])
enhancer.load_state_dict(torch.load(nn_model_path, map_location=device))
enhancer.to(device)
enhancer.eval()
if mfcc1.input_step == 'wave':
input_args = AR.filter_args(**kwargs)
reader = AR(input_path, **input_args)
else:
input_args = DRF.filter_args(**kwargs)
reader = DRF.create(input_path, **input_args)
writer = DWF.create(output_path, scp_sep=' ',
compress=compress,
compression_method=compression_method)
if write_num_frames is not None:
f_num_frames = open(write_num_frames, 'w')
for data in reader:
if mfcc1.input_step == 'wave':
key, x, fs = data
else:
key, x = data
logging.info('Extracting filter-banks for %s' % (key))
t1 = time.time()
y = mfcc1.compute(x)
# separate logE and filterbanks
logE = y[:,0]
y = y[:,1:]
#estimate log energy from filterbanks
logEy1 = logsumexp(y, axis=-1)
#we apply dummy identity network to fb
logging.info('Running enhancement network')
y = mvn.normalize(y)
y = apply_nnet(y, enhancer, chunk_size, context, device)
#lets rescale the logE based on enhanced filterbanks
logEy2 = logsumexp(y, axis=-1)
logE = logE + (logEy2 - logEy1)
# concatenate logE and filterbanks
y = np.concatenate((logE[:,None], y), axis=-1)
#apply DCT
logging.info('Applying DCT')
y = mfcc2.compute(y)
dt = (time.time() - t1)*1000
rtf = mfcc1.frame_shift*y.shape[0]/dt
logging.info('Extracted filter-banks for %s num-frames=%d elapsed-time=%.2f ms. real-time-factor=%.2f' %
(key, y.shape[0], dt, rtf))
writer.write([key], [y])
if write_num_frames is not None:
f_num_frames.write('%s %d\n' % (key, y.shape[0]))
mfcc1.reset()
if write_num_frames is not None:
f_num_frames.close()
y = np.concatenate((logE[:,None], y), axis=-1)
#apply DCT
logging.info('Applying DCT')
y = mfcc2.compute(y)
dt = (time.time() - t1)*1000
rtf = mfcc1.frame_shift*y.shape[0]/dt
logging.info('Extracted filter-banks for %s num-frames=%d elapsed-time=%.2f ms. real-time-factor=%.2f' %
(key, y.shape[0], dt, rtf))
writer.write([key], [y])
if write_num_frames is not None:
f_num_frames.write('%s %d\n' % (key, y.shape[0]))
mfcc1.reset()
if write_num_frames is not None:
f_num_frames.close()
if __name__ == "__main__":
parser=argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
fromfile_prefix_chars='@',
description='Compute filter-bank features and enhance with pytorch model')
parser.add_argument('--input', dest='input_path', required=True)
parser.add_argument('--output', dest='output_path', required=True)
parser.add_argument('--write-num-frames', dest='write_num_frames', default=None)