def run(args):
stft_kwargs = {
"frame_length": args.frame_length,
"frame_shift": args.frame_shift,
"window": args.window,
"center": args.center,
"transpose": True # F x T instead of T x F
}
spectrogram_reader = SpectrogramReader(args.wav_scp, **stft_kwargs)
for key, spectrogram in spectrogram_reader:
logger.info("Processing utterance {}...".format(key))
separated = auxiva(spectrogram, args.epochs)
for idx in range(separated.shape[0]):
samps = istft(
separated[idx],
**stft_kwargs,
norm=spectrogram_reader.samp_norm(key))
write_wav(
os.path.join(args.dst_dir, "{}.SRC{:d}.wav".format(
key, idx + 1)),
samps,
fs=args.fs)
logger.info("Processed {:d} utterances".format(len(spectrogram_reader)))
def run(args):
# shape: T x F, complex
stft_kwargs = {
"frame_length": args.frame_length,
"frame_shift": args.frame_shift,
"window": args.window,
"center": args.center,
}
spectrogram_reader = SpectrogramReader(args.wav_scp, **stft_kwargs)
mask_reader = NumpyReader(args.mask_scp) if args.numpy else ScriptReader(
args.mask_scp)
num_done = 0
with WaveWriter(args.dst_dir, fs=args.samp_freq) as writer:
for key, specs in spectrogram_reader:
if key in mask_reader:
num_done += 1
mask = mask_reader[key]
if args.transpose:
mask = np.transpose(mask)
logger.info("Processing utterance {}...".format(key))
if mask.shape != specs.shape:
raise ValueError(
"Dimention mismatch between mask and spectrogram"
"({0[0]} x {0[1]} vs {1[0]} x {1[1]}), need check configures"
.format(mask.shape, specs.shape))
nsamps = spectrogram_reader.nsamps(
key) if args.keep_length else None
norm = spectrogram_reader.samp_norm(key)
samps = istft(
specs * mask, **stft_kwargs, norm=norm, nsamps=nsamps)
writer.write(key, samps)
logger.info("Processed {:d} utterances over {:d}".format(
num_done, len(spectrogram_reader)))
def run(args):
# shape: T x F, complex
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center,
}
spectrogram_reader = SpectrogramReader(
args.wav_scp,
**stft_kwargs,
round_power_of_two=args.round_power_of_two)
phase_reader = None
if args.phase_ref:
phase_reader = SpectrogramReader(
args.phase_ref,
**stft_kwargs,
round_power_of_two=args.round_power_of_two)
logger.info("Using phase reference from {}".format(args.phase_ref))
MaskReader = {"numpy": NumpyReader, "kaldi": ScriptReader}
mask_reader = MaskReader[args.fmt](args.mask_scp)
num_done = 0
with WaveWriter(args.dst_dir, fs=args.samp_freq) as writer:
for key, specs in spectrogram_reader:
# specs: T x F
if key in mask_reader:
num_done += 1
mask = mask_reader[key]
# mask sure mask in T x F
_, F = specs.shape
if mask.shape[0] == F:
mask = np.transpose(mask)
logger.info("Processing utterance {}...".format(key))
if mask.shape != specs.shape:
raise ValueError(
"Dimention mismatch between mask and spectrogram"
"({0[0]} x {0[1]} vs {1[0]} x {1[1]}), need check configures"
.format(mask.shape, specs.shape))
nsamps = spectrogram_reader.nsamps(
key) if args.keep_length else None
norm = spectrogram_reader.samp_norm(key)
# use phase from ref
if phase_reader is not None:
angle = np.angle(phase_reader[key])
phase = np.exp(angle * 1j)
samps = istft(np.abs(specs) * mask * phase,
**stft_kwargs,
norm=norm,
nsamps=nsamps)
else:
samps = istft(specs * mask,
**stft_kwargs,
norm=norm,
nsamps=nsamps)
writer.write(key, samps)
logger.info("Processed {:d} utterances over {:d}".format(
num_done, len(spectrogram_reader)))
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": False # F x T
}
spectrogram_reader = SpectrogramReader(args.wav_scp, **stft_kwargs)
mask_reader = NumpyReader(args.mask_scp) if args.numpy else ScriptReader(
args.mask_scp)
num_bins = nfft(args.frame_length) // 2 + 1
supported_beamformer = {
"mvdr": MvdrBeamformer(num_bins),
"gevd": GevdBeamformer(num_bins),
"pmwf": PmwfBeamformer(num_bins)
}
beamformer = supported_beamformer[args.beamformer]
num_utts = 0
for key, stft_mat in spectrogram_reader:
if key in mask_reader:
num_utts += 1
norm = spectrogram_reader.samp_norm(key)
logger.info("Processing utterance {}(norm to {:.2f})...".format(
key, norm))
# prefer T x F
speech_mask = mask_reader[key]
if args.trans:
speech_mask = np.transpose(speech_mask)
# stft_enh, stft_mat: F x T
stft_enh = beamformer.run(speech_mask,
stft_mat,
normalize=args.postf)
# masking beamformer output if necessary
if args.mask:
stft_enh = stft_enh * np.transpose(speech_mask)
istft(os.path.join(args.dst_dir, '{}.wav'.format(key)),
stft_enh,
norm=norm,
fs=args.samp_freq,
**stft_kwargs)
logger.info("Processed {:d} utterances out of {:d}".format(
num_utts, len(spectrogram_reader)))
