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_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_utts = 0
fs = args.samp_freq,
for key, specs in spectrogram_reader:
if key in mask_reader:
num_utts += 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
istft(os.path.join(args.dst_dir, "{}.wav".format(key)),
specs * mask,
**stft_kwargs,
fs=fs,
nsamps=nsamps)
logger.info("Processed {} utterances".format(num_utts))
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):
# return complex result
stft_kwargs = {
"frame_length": args.frame_length,
"frame_shift": args.frame_shift,
"window": args.window,
"center": args.center
}
logger.info("Using mask: {}".format(args.mask.upper()))
mixture_reader = SpectrogramReader(
args.mix_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
ref_scp_list = args.ref_scp.split(",")
logger.info("Number of speakers: {:d}".format(len(ref_scp_list)))
targets_reader = [
SpectrogramReader(scp, **stft_kwargs) for scp in ref_scp_list
]
num_utts = 0
for key, mixture in tqdm(mixture_reader):
nsamps = mixture_reader.nsamps(key) if args.keep_length else None
skip = False
for reader in targets_reader:
if key not in reader:
logger.info("Skip utterance {}, missing targets".format(key))
skip = True
break
if skip:
continue
num_utts += 1
targets_list = [reader[key] for reader in targets_reader]
spk_masks = compute_mask(mixture, targets_list, args.mask)
for index, mask in enumerate(spk_masks):
samps = istft(mixture * mask, **stft_kwargs, nsamps=nsamps)
write_wav(os.path.join(args.dump_dir,
"spk{:d}/{}.wav".format(index + 1, key)),
samps,
fs=args.fs)
logger.info("Processed {} utterance!".format(num_utts))
def run(args):
# return complex result
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center
}
logger.info(f"Using mask: {args.mask.upper()}")
mixture_reader = SpectrogramReader(
args.mix_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
ref_scp_list = args.ref_scp.split(",")
logger.info(f"Number of speakers: {len(ref_scp_list)}")
targets_reader = [
SpectrogramReader(scp, **stft_kwargs) for scp in ref_scp_list
]
num_utts = 0
for key, mixture in tqdm(mixture_reader):
nsamps = mixture_reader.nsamps(key) if args.keep_length else None
skip = False
for reader in targets_reader:
if key not in reader:
logger.info(f"Skip utterance {key}, missing targets")
skip = True
break
if skip:
continue
num_utts += 1
targets_list = [reader[key] for reader in targets_reader]
spk_masks = compute_mask(mixture, targets_list, args.mask)
for index, mask in enumerate(spk_masks):
samps = inverse_stft(mixture * mask, **stft_kwargs, nsamps=nsamps)
write_wav(os.path.join(args.dump_dir, f"spk{index + 1}/{key}.wav"),
samps,
sr=args.sr)
logger.info(f"Processed {num_utts} utterance")
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(f"Using phase reference from {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.sf) as writer:
for key, specs in spectrogram_reader:
# if multi-channel, choose ch0
if specs.ndim == 3:
specs = specs[0]
# 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(f"Processing utterance {key}...")
if mask.shape != specs.shape:
raise ValueError(
"Dimention mismatch between mask and spectrogram" +
f"({mask.shape[0]} x {mask.shape[1]} vs " +
f"{specs.shape[0]} x {specs.shape[1]}), need " +
"check configures")
nsamps = spectrogram_reader.nsamps(
key) if args.keep_length else None
norm = spectrogram_reader.maxabs(
key) if args.mixed_norm else None
# use phase from ref
if phase_reader is not None:
angle = np.angle(phase_reader[key])
phase = np.exp(angle * 1j)
samps = inverse_stft(np.abs(specs) * mask * phase,
**stft_kwargs,
norm=norm,
nsamps=nsamps)
else:
samps = inverse_stft(specs * mask,
**stft_kwargs,
norm=norm,
nsamps=nsamps)
writer.write(key, samps)
logger.info(
f"Processed {num_done:d} utterances over {len(spectrogram_reader):d}")
