def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center,
"transpose": False
}
spectrogram_reader = SpectrogramReader(
args.wav_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
# F x N or B x F x N
weights = np.load(args.weights)
if weights.ndim == 2:
beamformer = FixedBeamformer(weights)
beam_index = None
else:
beamformer = [FixedBeamformer(w) for w in weights]
if not args.beam:
raise RuntimeError(
"--beam must be assigned, as there are multiple beams")
beam_index = ScpReader(args.beam, value_processor=lambda x: int)
with WaveWriter(args.dst_dir) as writer:
for key, stft_mat in spectrogram_reader:
logger.info(f"Processing utterance {key}...")
if beamformer:
beam = beam_index[key]
stft_enh = beamformer[beam].run(stft_mat)
else:
stft_enh = beamformer.run(stft_mat)
norm = spectrogram_reader.maxabs(key)
samps = inverse_stft(stft_enh, **stft_kwargs, norm=norm)
writer.write(key, samps)
logger.info(f"Processed {len(spectrogram_reader):d} utterances")
def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center,
"transpose": False
}
if args.geometry == "linear":
topo = str2tuple(args.linear_topo)
beamformer = LinearSDBeamformer(topo)
logger.info(f"Initialize LinearSDBeamformer for array: {topo}")
else:
beamformer = CircularSDBeamformer(args.circular_radius,
args.circular_around,
center=args.circular_center)
logger.info(
"Initialize CircularSDBeamformer for " +
f"radius = {args.circular_radius}, center = {args.circular_center}"
)
utt2doa = None
doa = None
if args.utt2doa:
utt2doa = ScpReader(args.utt2doa, value_processor=lambda x: float(x))
logger.info(f"Use --utt2doa={args.utt2doa} for each utterance")
else:
doa = args.doa
if not check_doa(args.geometry, doa):
logger.info(f"Invalid doa {doa:.2f} for {args.geometry} array")
logger.info(f"Use --doa={doa:.2f} for all utterances")
spectrogram_reader = SpectrogramReader(
args.wav_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
done = 0
with WaveWriter(args.dst_dir, sr=args.sr) as writer:
for key, stft_src in spectrogram_reader:
if utt2doa:
if key not in utt2doa:
continue
doa = utt2doa[key]
if not check_doa(args.geometry, doa):
logger.info(f"Invalid DoA {doa:.2f} for utterance {key}")
continue
stft_enh = beamformer.run(doa, stft_src, c=args.speed, sr=args.sr)
done += 1
norm = spectrogram_reader.maxabs(key)
samps = inverse_stft(stft_enh, **stft_kwargs, norm=norm)
writer.write(key, samps)
logger.info(f"Processed {done} utterances over {len(spectrogram_reader)}")
def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"round_power_of_two": args.round_power_of_two,
"window": args.window,
"center": args.center, # false to comparable with kaldi
"transpose": False # F x T
}
stft_reader = SpectrogramReader(args.wav_scp, **stft_kwargs)
if args.utt2idx:
utt2idx = ScpReader(args.utt2idx, value_processor=int)
logger.info(f"Using --utt2idx={args.utt2idx}")
else:
utt2idx = None
logger.info(f"Using --doa-idx={args.doa_idx}")
df_pair = [tuple(map(int, p.split(","))) for p in args.df_pair.split(";")]
if not len(df_pair):
raise RuntimeError(f"Bad configurations with --pair {args.pair}")
logger.info(f"Compute directional feature with {df_pair}")
# A x M x F
steer_vector = np.load(args.steer_vector)
num_done = 0
with ArchiveWriter(args.dup_ark, args.scp) as writer:
for key, stft in stft_reader:
# sv: M x F
if utt2idx is None:
idx = [int(v) for v in args.doa_idx.split(",")]
dfs = [
directional_feats(stft, steer_vector[i], df_pair=df_pair)
for i in idx
]
if len(dfs) == 1:
df = dfs[0]
else:
# N x T x F
dfs = np.stack(dfs)
df = dfs.transpose(1, 0, 2).reshape(dfs.shape[1], -1)
elif key in utt2idx:
# stft: M x F x T
df = directional_feats(stft,
steer_vector[utt2idx[key]],
df_pair=df_pair)
else:
logger.warn(f"Missing utt2idx for utterance {key}")
continue
writer.write(key, df)
num_done += 1
if not num_done % 1000:
logger.info(f"Processed {num_done:d} utterance...")
logger.info(f"Processed {num_done:d} utterances over {len(stft_reader):d}")
def parse_doa(args, online):
if args.utt2doa:
reader = ScpReader(
args.utt2doa,
value_processor=lambda doa: process_doa(doa, online),
num_tokens=-1)
utt2doa = reader.get
logger.info(f"Use --utt2doa={args.utt2doa} for each utterance")
else:
doa = process_doa(args.doa, online)
utt2doa = lambda _: doa
logger.info(f"Use --doa={doa:.2f} for all utterances")
return utt2doa
def run(args):
stft_kwargs = {
"frame_len": args.frame_len,
"frame_hop": args.frame_hop,
"window": args.window,
"center": args.center,
"transpose": False
}
utt2doa = None
doa = None
if args.utt2doa:
utt2doa = ScpReader(args.utt2doa, value_processor=lambda x: float(x))
logger.info(f"Use utt2doa {args.utt2doa} for each utterance")
else:
doa = args.doa
if doa < 0:
doa = 180 + doa
if doa < 0 or doa > 180:
raise RuntimeError(f"Invalid doa {doa:.2f} for --doa")
logger.info(f"Use DoA {doa:.2f} for all utterances")
spectrogram_reader = SpectrogramReader(
args.wav_scp,
round_power_of_two=args.round_power_of_two,
**stft_kwargs)
done = 0
topo = str2tuple(args.linear_topo)
beamformer = LinearDSBeamformer(topo)
logger.info(f"Initialize channel LinearDSBeamformer for array: {topo}")
with WaveWriter(args.dst_dir, fs=args.fs) as writer:
for key, stft_src in spectrogram_reader:
if utt2doa:
if key not in utt2doa:
continue
doa = utt2doa[key]
if doa < 0:
doa = 180 + doa
if doa < 0 or doa > 180:
logger.info(f"Invalid doa {doa:.2f} for utterance {key}")
continue
stft_enh = beamformer.run(doa, stft_src, c=args.speed, sr=args.fs)
done += 1
norm = spectrogram_reader.maxabs(key)
samps = inverse_stft(stft_enh, **stft_kwargs, norm=norm)
writer.write(key, samps)
logger.info(f"Processed {done} utterances over {len(spectrogram_reader)}")
def __init__(self, spk2class=None):
self.s2c = ScpReader(spk2class) if spk2class else None
self.snr = defaultdict(float)
self.cnt = defaultdict(int)
def __init__(self, text):
self.text_reader = [
ScpReader(t, num_tokens=-1, restrict=False)
for t in text.split(",")
]
