def otf_utt_generator(data_triplets, rir, noise, args):
"""
Args:
data_lst: list of mrk and seq of input audios, and label ark
rir: list of rir, List[AudioSegment]
noise: list of noise, List[AudioSegment]
args: argumnets for loader
"""
max_len = args.max_len
batch_size = args.batch_size
data_buffer = np.zeros((batch_size, max_len, get_inputdim(args)),
dtype=np.float32)
target_buffer = np.zeros((batch_size, max_len), dtype=np.int32)
len_buffer = np.zeros(batch_size, dtype=np.int32)
ali_len = np.zeros(batch_size, dtype=np.int32)
batch_idx = 0
valid_idx = 0
target_len = 0
batch_max_len = -1
target_max_len = -1
#rates for speed perturbation
speed_rate = [float(rate) for rate in args.speed_rate.split(',')]
#volume level perturbation
gain_lo, gain_hi = [-float(gain) for gain in args.gain_range.split(',')]
#snr range for noise perturbation: 0-20db with mean of 10
#mu, sigma = 10, 10
#lo, hi = (0 - mu) / sigma, (20 - mu) / sigma
#Fbank config
po = ParseOptions('')
fbank_opt = FbankOptions()
fbank_opt.register(po)
#fbank_opt = MfccOptions()
#fbank_opt.register(po)
po.read_config_file(args.feat_config)
fbank = Fbank(fbank_opt)
#fbank = Mfcc(fbank_opt)
for data_triplet in data_triplets:
mrk_fn, seq_fn = data_triplet[0], data_triplet[1]
ali_rspec = data_triplet[2]
with open(mrk_fn, 'r', encoding='utf-8') as mrk,\
open(seq_fn, 'rb') as seq:
ali_reader = SequentialIntVectorReader(ali_rspec)
for line, (uttid1, ali) in zip(mrk, ali_reader):
uttid = line.split()[0]
assert uttid == uttid1
seq.seek(int(line.split()[1]))
num_bytes = int(line.split()[2])
num_bytes -= num_bytes % 2
audio_bytes = seq.read(num_bytes)
audio_np = np.frombuffer(audio_bytes, dtype='int16')
#data augmentation function goes here
audio_seg = AudioSegment(audio_np, args.sample_rate)
#speed perturbation
spr = speed_rate[randint(0, len(speed_rate) - 1)]
audio_seg.change_speed(spr)
audio_seg.normalize(np.random.uniform(gain_lo, gain_hi))
#noise adding example:
#snr = truncnorm.rvs(lo, hi, scale=sigma, loc=mu, size=1)
#audio_seg.add_noise(noise[randint(0, len(noise)-1)], snr)
#rir adding example:
#audio_seg.convolve_and_normalize(rir[randint(0, len(rir)-1)])
audio_np = audio_seg._convert_samples_from_float32(\
audio_seg.samples, 'int16')
wave_1ch = Vector(audio_np)
feats = fbank.compute_features(wave_1ch,
args.sample_rate,
vtnl_warp=1.0)
ali = np.array(ali)
if args.reverse_labels:
ali = ali[::-1]
if args.SOS >= 0:
ali = np.concatenate(([args.SOS], ali))
if args.EOS >= 0:
ali = np.concatenate((ali, [args.EOS]))
feats = _matrix_ext.matrix_to_numpy(feats)
utt_len = feats.shape[0] // args.stride + \
int(feats.shape[0] % args.stride != 0)
#limits on T*U products due to RNNT.
#this is pretty hacky now
if ali.shape[0] * utt_len // 3 <= args.TU_limit:
ali_len[valid_idx] = ali.shape[0]
data_buffer[valid_idx, :utt_len, :] = \
splice(feats, args.lctx, args.rctx)[::args.stride]
target_buffer[valid_idx, :ali_len[valid_idx]] = ali
len_buffer[valid_idx] = utt_len
if utt_len > batch_max_len:
batch_max_len = utt_len
if ali_len[valid_idx] > target_max_len:
target_max_len = ali_len[valid_idx]
valid_idx += 1
batch_idx += 1
if batch_idx == batch_size:
for b in range(valid_idx):
utt_len = len_buffer[b]
target_len = ali_len[b]
#data and target padding
if utt_len > 0:
data_buffer[b, utt_len:batch_max_len, :] = \
data_buffer[b, utt_len-1, :]
target_buffer[b, target_len:target_max_len] = \
args.padding_tgt
data = data_buffer[:valid_idx, :batch_max_len, :]
target = target_buffer[:valid_idx, :target_max_len]
if not args.batch_first:
data = np.transpose(data, (1, 0, 2))
target = np.transpose(target, (1, 0))
data = torch.from_numpy(np.copy(data))
target = torch.from_numpy(np.copy(target))
lens = torch.from_numpy(np.copy(len_buffer[:valid_idx]))
ali_lens = torch.from_numpy(np.copy(ali_len[:valid_idx]))
if valid_idx > 0:
#not doing cuda() here, in main process instead
yield data, target, lens, ali_lens
else:
yield None, None, \
torch.IntTensor([0]), torch.IntTensor([0])
batch_idx = 0
valid_idx = 0
target_len = 0
batch_max_len = -1
target_max_len = -1
ali_reader.close()
yield None
mrk_fn = line.split()[0]
seq_fn = line.split()[1]
with open(mrk_fn, 'r', encoding='utf-8') as mrk, \
open(seq_fn, 'rb') as seq:
for mrk_line in mrk:
seq.seek(int(mrk_line.split()[1]))
num_bytes = int(mrk_line.split()[2])
#this is making sure even number of bytes
num_bytes -= num_bytes % 2
audio_bytes = seq.read(num_bytes)
audio_np = np.frombuffer(audio_bytes, dtype='int16')
audio_seg = AudioSegment(audio_np, args.sample_rate)
spr = speed_rate[randint(0, len(speed_rate) - 1)]
audio_seg.change_speed(spr)
#-55 to -10 db
audio_seg.normalize(np.random.uniform(-55, -10))
audio_np = audio_seg._convert_samples_from_float32(\
audio_seg.samples, 'int16')
wave_1ch = Vector(audio_np)
feats = fbank.compute_features(wave_1ch,
args.sample_rate,
vtnl_warp=1.0)
if args.cmn:
feats = _matrix_ext.matrix_to_numpy(feats)
feats -= np.mean(feats, axis=0)
feats = Matrix(feats)
cmvn.accumulate(feats)
cmvn.write_stats(args.cmvn_stats, binary=False)