def __init__(self, rt60_opt, absc_opt, room_dim):
"""
rt60_opt: "" or "a,b", higher priority than absc_opt
absc_opt: tuple like (a,b)
room_dim: str like "a,b;c,d;e,d"
"""
self.rt60_opt = rt60_opt
if not rt60_opt:
self.absc = UniformSampler(absc_opt)
else:
rt60_r = str2tuple(rt60_opt)
self.rt60 = UniformSampler(rt60_r)
dim_range = [str2tuple(t) for t in room_dim.split(";")]
if len(dim_range) != 3:
raise RuntimeError(f"Wrong format with --room-dim={room_dim}")
self.dim_sampler = [UniformSampler(c) for c in dim_range]
def run(args):
if args.geometry == "linear":
topo = np.array(str2tuple(args.linear_topo))
candidate_doa = np.linspace(0, 180, args.num_doas)
else:
topo = None
step = 360 / args.num_doas
candidate_doa = np.arange(0, 360, step)
sv = []
for doa in candidate_doa:
if topo is None:
sv.append(
circular_steer_vector(args.circular_radius,
args.circular_around,
doa,
args.num_bins,
c=args.speed,
sr=args.sr,
center=args.circular_center))
else:
sv.append(
linear_steer_vector(topo,
doa,
args.num_bins,
c=args.speed,
sr=args.sr))
# A x F x M
sv = np.stack(sv)
# norm or not
if args.normalize:
sv = sv / sv.shape[-1]**0.5
# A x M x F
sv = sv.transpose(0, 2, 1)
np.save(args.steer_vector, sv)
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 __init__(self, args):
if args.gpu and not gpu_rir_available:
raise RuntimeError("Please install gpuRIR first if --gpu=True")
# make dump dir
Path(args.dump_dir).mkdir(exist_ok=True, parents=True)
self.rirs_cfg = []
self.room_generator = RoomGenerator(args.rt60, args.abs_range,
args.room_dim)
self.mx, self.my = args.array_relx, args.array_rely
self.array_topo = [str2tuple(t) for t in args.array_topo.split(";")]
self.sr = args.sample_rate
self.args = args
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 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,
"transpose": True
}
steer_vector = np.load(args.steer_vector)
logger.info(f"Shape of the steer vector: {steer_vector.shape}")
num_doa, _, _ = steer_vector.shape
min_doa, max_doa = str2tuple(args.doa_range)
if args.output == "radian":
angles = np.linspace(min_doa * np.pi / 180, max_doa * np.pi / 180,
num_doa + 1)
else:
angles = np.linspace(min_doa, max_doa, num_doa + 1)
spectrogram_reader = SpectrogramReader(args.wav_scp, **stft_kwargs)
mask_reader = None
if args.mask_scp:
mask_reader = [NumpyReader(scp) for scp in args.mask_scp.split(",")]
online = (args.chunk_len > 0 and args.look_back > 0)
if online:
logger.info("Set up in online mode: chunk_len " +
f"= {args.chunk_len}, look_back = {args.look_back}")
if args.backend == "srp":
split_index = lambda sstr: [
tuple(map(int, p.split(","))) for p in sstr.split(";")
]
srp_pair = split_index(args.srp_pair)
srp_pair = ([t[0] for t in srp_pair], [t[1] for t in srp_pair])
logger.info(f"Choose srp-based algorithm, srp pair is {srp_pair}")
else:
srp_pair = None
with open(args.doa_scp, "w") as doa_out:
for key, stft in spectrogram_reader:
# stft: M x T x F
_, _, F = stft.shape
if mask_reader:
# T x F => F x T
mask = [r[key] for r in mask_reader] if mask_reader else None
if args.mask_eps >= 0 and len(mask_reader) > 1:
mask = add_wta(mask, eps=args.mask_eps)
mask = mask[0]
# F x T => T x F
if mask.shape[-1] != F:
mask = mask.transpose()
else:
mask = None
if not online:
if srp_pair:
idx = srp_ssl(stft,
steer_vector,
srp_pair=srp_pair,
mask=mask)
else:
idx = ml_ssl(stft,
steer_vector,
mask=mask,
compression=-1,
eps=EPSILON)
doa = angles[idx]
logger.info(f"Processing utterance {key}: {doa:.4f}")
doa_out.write(f"{key}\t{doa:.4f}\n")
else:
logger.info(f"Processing utterance {key}...")
_, T, _ = stft.shape
online_doa = []
for t in range(0, T, args.chunk_len):
s = max(t - args.look_back, 0)
if mask is not None:
chunk_mask = mask[..., s:t + args.chunk_len]
else:
chunk_mask = None
stft_chunk = stft[:, s:t + args.chunk_len, :]
if srp_pair:
idx = srp_ssl(stft_chunk,
steer_vector,
srp_pair=srp_pair,
mask=chunk_mask)
else:
idx = ml_ssl(stft_chunk,
steer_vector,
mask=chunk_mask,
compression=-1,
eps=EPSILON)
doa = angles[idx]
online_doa.append(doa)
doa_str = " ".join([f"{d:.4f}" for d in online_doa])
doa_out.write(f"{key}\t{doa_str}\n")
logger.info(f"Processing {len(spectrogram_reader)} utterance done")
