def run(args):
griffin_lim_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center,
"transpose": True,
"epochs": args.epochs
}
feature_reader = ScriptReader(args.feat_scp)
if args.fbank:
mel_kwargs = {
"n_mels": args.num_bins,
"fmin": args.min_freq,
"fmax": args.max_freq,
"htk": True
}
# N x F
mel_weights = audio_lib.filters.mel(args.samp_freq,
nfft(args.frame_length),
**mel_kwargs)
# F x N
mel_inv_weights = np.linalg.pinv(mel_weights)
with WaveWriter(
args.dump_dir, fs=args.samp_freq,
normalize=args.normalize) as writer:
for key, spec in feature_reader:
# if log, tranform to linear
if args.apply_log:
spec = np.exp(spec)
# convert fbank to spectrum
# feat: T x N
if args.fbank:
spec = np.maximum(spec @ np.transpose(mel_inv_weights),
EPSILON)
# if power spectrum, tranform to magnitude spectrum
if args.apply_pow:
spec = np.sqrt(spec)
if spec.shape[1] - 1 != nfft(args.frame_length) // 2:
raise RuntimeError("Seems missing --fbank options?")
# griffin lim
samps = griffin_lim(spec, **griffin_lim_kwargs)
writer.write(key, samps)
logger.info("Processed {:d} utterance done".format(len(feature_reader)))
def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center,
}
FeatureReader = {"numpy": NumpyReader, "kaldi": ScriptReader}
feature_reader = FeatureReader[args.fmt](args.feat_scp)
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}")
with WaveWriter(args.dump_dir, fs=args.sr,
normalize=args.normalize) as writer:
for key, spec in feature_reader:
logger.info(f"Processing utterance {key}...")
# if log, tranform to linear
if args.apply_log:
spec = np.exp(spec)
# if power spectrum, tranform to magnitude spectrum
if args.apply_pow:
spec = np.sqrt(spec)
if phase_reader is None:
# griffin lim
samps = griffin_lim(spec,
epoches=args.epoches,
transpose=True,
norm=0.8,
**stft_kwargs)
else:
if key not in phase_reader:
raise KeyError(f"Missing key {key} in phase reader")
ref = phase_reader[key]
angle = np.angle(ref[0] if ref.ndim == 3 else ref)
phase = np.exp(angle * 1j)
samps = inverse_stft(spec * phase, **stft_kwargs, norm=0.8)
writer.write(key, samps)
logger.info(f"Processed {len(feature_reader)} utterance done")