speech_recognition/FunASR
1
0
Fork 0
mirror of https://github.com/modelscope/FunASR synced 2025-09-15 14:48:36 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix bug for onnx paraformer-long

Browse Source
This commit is contained in:
shixian.shi 2023-03-20 14:29:37 +08:00
parent 977519f54b
commit fc6968b44a
2 changed files with 12 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
funasr/runtime/python/onnxruntime/rapid_paraformer/paraformer_onnx.py
View File

@ -75,14 +75,17 @@ class Paraformer():
preds = self.decode(am_scores, valid_token_lens)
if us_cif_peak is None:
for pred in preds:
pred = sentence_postprocess(pred)
asr_res.append({'preds': pred})
else:
for pred, us_cif_peak_ in zip(preds, us_cif_peak):
text, tokens = pred
timestamp, timestamp_total = time_stamp_lfr6_onnx(us_cif_peak_, copy.copy(tokens))
raw_tokens = pred
timestamp, timestamp_raw = time_stamp_lfr6_onnx(us_cif_peak_, copy.copy(raw_tokens))
text_proc, timestamp_proc, _ = sentence_postprocess(raw_tokens, timestamp_raw)
# logging.warning(timestamp)
if len(self.plot_timestamp_to):
self.plot_wave_timestamp(waveform_list[0], timestamp_total, self.plot_timestamp_to)
asr_res.append({'preds': text, 'timestamp': timestamp})
self.plot_wave_timestamp(waveform_list[0], timestamp, self.plot_timestamp_to)
asr_res.append({'preds': text_proc, 'timestamp': timestamp_proc, "raw_tokens": raw_tokens})
return asr_res
def plot_wave_timestamp(self, wav, text_timestamp, dest):
@ -181,6 +184,6 @@ class Paraformer():
# Change integer-ids to tokens
token = self.converter.ids2tokens(token_int)
token = token[:valid_token_num-self.pred_bias]
texts = sentence_postprocess(token)
return texts
# texts = sentence_postprocess(token)
return token

6
funasr/runtime/python/onnxruntime/rapid_paraformer/utils/timestamp_utils.py
View File

@ -48,12 +48,12 @@ def time_stamp_lfr6_onnx(us_cif_peak, char_list, begin_time=0.0, total_offset=-1
timestamp_list[i][0] = timestamp_list[i][0] + begin_time / 1000.0
timestamp_list[i][1] = timestamp_list[i][1] + begin_time / 1000.0
assert len(new_char_list) == len(timestamp_list)
res_total = []
res_str = ""
for char, timestamp in zip(new_char_list, timestamp_list):
res_total.append([char, timestamp[0], timestamp[1]]) # += "{} {} {};".format(char, timestamp[0], timestamp[1])
res_str += "{} {} {};".format(char, timestamp[0], timestamp[1])
res = []
for char, timestamp in zip(new_char_list, timestamp_list):
if char != '<sil>':
res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
return res, res_total
return res_str, res