def get_labformat(timestamp, subsample):
begin = 0
duration = 0
labformat = []
for idx, t in enumerate(timestamp):
# 25ms frame_length,10ms hop_length, 1/subsample
subsample = get_subsample(configs)
# time duration
duration = len(t) * 0.01 * subsample
if idx < len(timestamp) - 1:
print("{:.2f} {:.2f} {}".format(begin, begin + duration,
char_dict[t[-1]]))
labformat.append("{:.2f} {:.2f} {}\n".format(
begin, begin + duration, char_dict[t[-1]]))
else:
non_blank = 0
for i in t:
if i != 0:
token = i
break
print("{:.2f} {:.2f} {}".format(begin, begin + duration,
char_dict[token]))
labformat.append("{:.2f} {:.2f} {}\n".format(
begin, begin + duration, char_dict[token]))
begin = begin + duration
return labformat
mask = make_pad_mask(encoder_out_lens) # (B, maxlen)
topk_index = topk_index.masked_fill_(mask, eos) # (B, maxlen)
alignment = [hyp.tolist() for hyp in topk_index]
hyps = [remove_duplicates_and_blank(hyp) for hyp in alignment]
for index, i in enumerate(key):
content = []
if len(hyps[index]) > 0:
for w in hyps[index]:
if w == eos:
break
content.append(char_dict[w])
f_ctc_results.write('{} {}\n'.format(i, " ".join(content)))
f_ctc_results.flush()
for index, i in enumerate(key):
timestamp = get_frames_timestamp(alignment[index])
subsample = get_subsample(configs)
word_seq, word_time = get_labformat_frames(
timestamp, subsample, char_dict)
for index_j in range(len(word_seq)):
for keyword in word_unit_list:
keyword_len = len(word_unit_dict[keyword])
if index_j + keyword_len > len(word_seq):
continue
if (word_seq[index_j:index_j +
keyword_len] == word_unit_dict[keyword]):
f_keyword_results.write("{} {} {} {} {}\n".format(
word_id_dict[keyword], i,
word_time[index_j][0],
word_time[index_j + keyword_len - 1][1], 0.0))
f_keyword_results.flush()
f_keyword_results.close()