def __init__(self,
mix_scp: str = "",
doa_scp: Union[str, List[str]] = "",
emb_scp: str = "",
ref_scp: Union[str, List[str]] = "",
sr: int = 16000) -> None:
self.mix = AudioReader(mix_scp, sr=sr)
if isinstance(ref_scp, list):
self.ref = [AudioReader(ref, sr=sr) for ref in ref_scp]
self.num_ref = len(ref_scp)
elif ref_scp:
self.ref = AudioReader(ref_scp, sr=sr)
self.num_ref = 1
else:
self.ref = None
self.num_ref = 0
self.num_doa = 0
if isinstance(doa_scp, list):
self.doa = [
BaseReader(doa, value_processor=lambda x: np.float32(x))
for doa in doa_scp
]
self.num_doa = len(doa_scp)
elif not doa_scp:
self.doa = None
else:
self.doa = BaseReader(doa_scp,
value_processor=lambda x: np.float32(x))
self.num_doa = 1
self.emb = NumpyReader(emb_scp) if emb_scp else None
def _pre_process(self,
text: str,
utt2dur: str,
max_token_num: int = 400,
min_token_num: int = 2,
skip_utts: str = "",
max_dur: float = 3000,
min_dur: float = 40) -> List[Dict]:
"""
Preprocess function to filter the utterances
"""
if skip_utts:
with open(skip_utts, "r") as skip_fd:
skip_keys = [k.strip() for k in skip_fd.readlines()]
else:
skip_keys = []
utt2dur = BaseReader(utt2dur, value_processor=float)
if self.vocab_dict:
text_reader = BaseReader(text, num_tokens=-1, restrict=False)
else:
text_reader = BaseReader(
text,
value_processor=lambda tok: list(map(int, tok)),
num_tokens=-1,
restrict=False)
token_set = []
drop_utts = 0
for key, tokens in text_reader:
num_toks = len(tokens)
if num_toks > max_token_num or num_toks < min_token_num:
drop_utts += 1
continue
if key not in utt2dur:
drop_utts += 1
continue
if key in skip_keys:
continue
num_frames = utt2dur[key]
if num_frames < min_dur or num_frames > max_dur:
drop_utts += 1
continue
token_set.append({
"key": key,
"dur": num_frames,
"len": num_toks,
"tok": tokens
})
# long -> short
token_set = sorted(token_set, key=lambda d: d["dur"], reverse=True)
if drop_utts:
warnings.warn(f"Drop {drop_utts} utterances")
return token_set
def run(args):
scp_out = open(args.scp, "w") if args.scp else None
with open(args.wav_ark, "wb") as wav_ark:
reader = BaseReader(args.wav_scp, num_tokens=2, restrict=True)
done = 0
for key, value in reader:
wav_ark.write(str.encode(key + " "))
offset = wav_ark.tell()
if value[-1] == "|":
p = subprocess.Popen(value[:-1],
shell=True,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE)
[stdout, stderr] = p.communicate()
if p.returncode != 0:
stderr = bytes.decode(stderr)
raise RuntimeError(
f"Running command: \"{value[:-1]}\" failed: {stderr}")
wav_ark.write(stdout)
else:
with open(value, "rb") as wav:
wav_ark.write(wav.read())
if scp_out:
scp_out.write(f"{key}\t{args.wav_ark}:{offset}\n")
done += 1
if done % 200 == 0:
print(f"Processed {done} utterances...", flush=True)
print(f"Archive {len(reader)} utterances to {args.wav_ark}")
if scp_out:
scp_out.close()
def __init__(self,
spk2class: Optional[str] = None,
name: str = "UNK",
unit: str = "UNK") -> None:
self.s2c = BaseReader(spk2class) if spk2class else None
self.val = defaultdict(float)
self.name = name
self.unit = unit
def __init__(self,
simu_cfg: str,
return_in_egs: List[str] = ["mix"]) -> None:
self.simu_cfg = BaseReader(simu_cfg, num_tokens=-1)
self.parser = make_argparse()
self.return_in_egs = return_in_egs
def __init__(self, text_descriptor, cer=False):
self.readers = [
BaseReader(t, num_tokens=-1, restrict=False)
for t in text_descriptor.split(",")
]
self.cer = cer