def run(args):
stft_kwargs = {
"frame_length": args.frame_length,
"frame_shift": args.frame_shift,
"window": args.window,
"center": args.center, # false to comparable with kaldi
"transpose": True # T x F
}
wpe_kwargs = {
"taps": args.taps,
"delay": args.delay,
"iters": args.iters,
"psd_context": args.context
}
spectrogram_reader = SpectrogramReader(args.wav_scp, **stft_kwargs)
if not os.path.exists(args.dst_dir):
os.makedirs(args.dst_dir)
for key, reverbed in spectrogram_reader:
# N x T x F => F x N x T
reverbed = np.transpose(reverbed, [2, 0, 1])
# F x N x T
dereverb = wpe(reverbed, **wpe_kwargs)
# F x N x T => N x T x F
dereverb = np.transpose(dereverb, [1, 2, 0])
# write for each channel
for chid in range(dereverb.shape[0]):
samps = istft(dereverb[chid], **stft_kwargs)
write_wav(os.path.join(args.dst_dir,
"{}.CH{:d}.wav".format(key, chid + 1)),
samps,
fs=args.samp_freq)
logger.info("Processed {:d} utterances".format(len(spectrogram_reader)))
def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center, # false to comparable with kaldi
"transpose": True # T x F
}
wpe_kwargs = {
"num_iters": args.num_iters,
"context": args.context,
"taps": args.taps,
"delay": args.delay
}
spectrogram_reader = SpectrogramReader(
args.wav_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
num_done = 0
with WaveWriter(args.dst_dir, fs=args.sr) as writer:
for key, reverbed in spectrogram_reader:
logger.info(f"Processing utt {key}...")
# N x T x F => F x N x T
reverbed = np.transpose(reverbed, (2, 0, 1))
try:
if args.nara_wpe:
from nara_wpe.wpe import wpe_v8
# T x F x N
dereverb = wpe_v8(reverbed,
taps=args.taps,
delay=args.delay,
iterations=args.num_iters,
psd_context=args.context)
else:
dereverb = wpe(reverbed, **wpe_kwargs)
except np.linalg.LinAlgError:
logger.warn(f"{key}: Failed cause LinAlgError in wpe")
continue
# F x N x T => N x T x F
dereverb = np.transpose(dereverb, (1, 2, 0))
# dump multi-channel
samps = np.stack(
[inverse_stft(spectra, **stft_kwargs) for spectra in dereverb])
writer.write(key, samps)
# show progress cause slow speed
num_done += 1
if not num_done % 100:
logger.info(f"Processed {num_done:d} utterances...")
logger.info(
f"Processed {num_done:d} utterances over {len(spectrogram_reader):d}")
def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center, # false to comparable with kaldi
"transpose": True # T x F
}
wpe_kwargs = {
"num_iters": args.num_iters,
"context": args.context,
"taps": args.taps,
"delay": args.delay
}
spectrogram_reader = SpectrogramReader(
args.wav_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
num_done = 0
with WaveWriter(args.dst_dir, fs=args.samp_fs) as writer:
for key, reverbed in spectrogram_reader:
logger.info("Processing utt {}...".format(key))
# N x T x F => F x N x T
reverbed = np.transpose(reverbed, (2, 0, 1))
try:
# F x N x T
dereverb = wpe(reverbed, **wpe_kwargs)
except np.linalg.LinAlgError:
logger.warn("{}: Failed cause LinAlgError in wpe".format(key))
continue
# F x N x T => N x T x F
dereverb = np.transpose(dereverb, (1, 2, 0))
# dump multi-channel
samps = np.stack(
[istft(spectra, **stft_kwargs) for spectra in dereverb])
writer.write(key, samps)
# show progress cause slow speed
num_done += 1
if not num_done % 100:
logger.info("Processed {:d} utterances...".format(num_done))
logger.info("Processed {:d} utterances over {:d}".format(
num_done, len(spectrogram_reader)))