From 80e6c258cf89b5f11f4e52a4cc5a9cf2e95aa7be Mon Sep 17 00:00:00 2001
From: Yuekai Zhang <zhangyuekai@foxmail.com>
Date: 星期一, 06 三月 2023 16:48:02 +0800
Subject: [PATCH] update token list
---
funasr/utils/timestamp_tools.py | 189 +++++++++++++++++++++++++----------------------
1 files changed, 100 insertions(+), 89 deletions(-)
diff --git a/funasr/utils/timestamp_tools.py b/funasr/utils/timestamp_tools.py
index 3afaa40..4a367f8 100644
--- a/funasr/utils/timestamp_tools.py
+++ b/funasr/utils/timestamp_tools.py
@@ -4,96 +4,107 @@
import numpy as np
from typing import Any, List, Tuple, Union
-def cut_interval(alphas: torch.Tensor, start: int, end: int, tail: bool):
- if not tail:
- if end == start + 1:
- cut = (end + start) / 2.0
- else:
- alpha = alphas[start+1: end].tolist()
- reverse_steps = 1
- for reverse_alpha in alpha[::-1]:
- if reverse_alpha > 0.35:
- reverse_steps += 1
- else:
- break
- cut = end - reverse_steps
+
+def time_stamp_lfr6_pl(us_alphas, us_cif_peak, char_list, begin_time=0.0, end_time=None):
+ if not len(char_list):
+ return []
+ START_END_THRESHOLD = 5
+ TIME_RATE = 10.0 * 6 / 1000 / 3 # 3 times upsampled
+ if len(us_alphas.shape) == 3:
+ alphas, cif_peak = us_alphas[0], us_cif_peak[0] # support inference batch_size=1 only
else:
- if end != len(alphas) - 1:
- cut = end + 1
- else:
- cut = start + 1
- return float(cut)
-
-def time_stamp_lfr6(alphas: torch.Tensor, speech_lengths: torch.Tensor, raw_text: List[str], begin: int = 0, end: int = None):
- time_stamp_list = []
- alphas = alphas[0]
- text = copy.deepcopy(raw_text)
- if end is None:
- time = speech_lengths * 60 / 1000
- sacle_rate = (time / speech_lengths[0]).tolist()
+ alphas, cif_peak = us_alphas, us_cif_peak
+ num_frames = cif_peak.shape[0]
+ if char_list[-1] == '</s>':
+ char_list = char_list[:-1]
+ # char_list = [i for i in text]
+ timestamp_list = []
+ # for bicif model trained with large data, cif2 actually fires when a character starts
+ # so treat the frames between two peaks as the duration of the former token
+ fire_place = torch.where(cif_peak>1.0-1e-4)[0].cpu().numpy() - 1.5
+ num_peak = len(fire_place)
+ assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
+ # begin silence
+ if fire_place[0] > START_END_THRESHOLD:
+ char_list.insert(0, '<sil>')
+ timestamp_list.append([0.0, fire_place[0]*TIME_RATE])
+ # tokens timestamp
+ for i in range(len(fire_place)-1):
+ # the peak is always a little ahead of the start time
+ # timestamp_list.append([(fire_place[i]-1.2)*TIME_RATE, fire_place[i+1]*TIME_RATE])
+ timestamp_list.append([(fire_place[i])*TIME_RATE, fire_place[i+1]*TIME_RATE])
+ # cut the duration to token and sil of the 0-weight frames last long
+ # tail token and end silence
+ if num_frames - fire_place[-1] > START_END_THRESHOLD:
+ _end = (num_frames + fire_place[-1]) / 2
+ timestamp_list[-1][1] = _end*TIME_RATE
+ timestamp_list.append([_end*TIME_RATE, num_frames*TIME_RATE])
+ char_list.append("<sil>")
else:
- time = (end - begin) / 1000
- sacle_rate = (time / speech_lengths[0]).tolist()
-
- predictor = (alphas > 0.5).int()
- fire_places = torch.nonzero(predictor == 1).squeeze(1).tolist()
-
- cuts = []
- npeak = int(predictor.sum())
- nchar = len(raw_text)
- if npeak - 1 == nchar:
- fire_places = torch.where((alphas > 0.5) == 1)[0].tolist()
- for i in range(len(fire_places)):
- if fire_places[i] < len(alphas) - 1:
- if 0.05 < alphas[fire_places[i]+1] < 0.5:
- fire_places[i] += 1
- elif npeak < nchar:
- lost_num = nchar - npeak
- lost_fire = speech_lengths[0].tolist() - fire_places[-1]
- interval_distance = lost_fire // (lost_num + 1)
- for i in range(1, lost_num + 1):
- fire_places.append(fire_places[-1] + interval_distance)
- elif npeak - 1 > nchar:
- redundance_num = npeak - 1 - nchar
- for i in range(redundance_num):
- fire_places.pop()
-
- cuts.append(0)
- start_sil = True
- if start_sil:
- text.insert(0, '<sil>')
-
- for i in range(len(fire_places)-1):
- cuts.append(cut_interval(alphas, fire_places[i], fire_places[i+1], tail=(i==len(fire_places)-2)))
-
- for i in range(2, len(fire_places)-2):
- if fire_places[i-2] == fire_places[i-1] - 1 and fire_places[i-1] != fire_places[i] - 1:
- cuts[i-1] += 1
-
- if cuts[-1] != len(alphas) - 1:
- text.append('<sil>')
- cuts.append(speech_lengths[0].tolist())
- cuts.insert(-1, (cuts[-1] + cuts[-2]) * 0.5)
- sec_fire_places = np.array(cuts) * sacle_rate
- for i in range(1, len(sec_fire_places) - 1):
- start, end = sec_fire_places[i], sec_fire_places[i+1]
- if i == len(sec_fire_places) - 2:
- end = time
- time_stamp_list.append([int(round(start, 2) * 1000) + begin, int(round(end, 2) * 1000) + begin])
- text = text[1:]
- if npeak - 1 == nchar or npeak > nchar:
- return time_stamp_list[:-1]
- else:
- return time_stamp_list
+ timestamp_list[-1][1] = num_frames*TIME_RATE
+ if begin_time: # add offset time in model with vad
+ for i in range(len(timestamp_list)):
+ timestamp_list[i][0] = timestamp_list[i][0] + begin_time / 1000.0
+ timestamp_list[i][1] = timestamp_list[i][1] + begin_time / 1000.0
+ res_txt = ""
+ for char, timestamp in zip(char_list, timestamp_list):
+ res_txt += "{} {} {};".format(char, timestamp[0], timestamp[1])
+ res = []
+ for char, timestamp in zip(char_list, timestamp_list):
+ if char != '<sil>':
+ res.append([int(timestamp[0] * 1000), int(timestamp[1] * 1000)])
+ return res
-def time_stamp_lfr6_advance(tst: List, text: str):
- # advanced timestamp prediction for BiCIF_Paraformer using upsampled alphas
- ds_alphas, ds_cif_peak, us_alphas, us_cif_peak = tst
- if text.endswith('</s>'):
- text = text[:-4]
- else:
- text = text[:-1]
- logging.warning("found text does not end with </s>")
- assert int(ds_alphas.sum() + 1e-4) - 1 == len(text)
-
+def time_stamp_sentence(punc_id_list, time_stamp_postprocessed, text_postprocessed):
+ res = []
+ if text_postprocessed is None:
+ return res
+ if time_stamp_postprocessed is None:
+ return res
+ if len(time_stamp_postprocessed) == 0:
+ return res
+ if len(text_postprocessed) == 0:
+ return res
+ if punc_id_list is None or len(punc_id_list) == 0:
+ res.append({
+ 'text': text_postprocessed.split(),
+ "start": time_stamp_postprocessed[0][0],
+ "end": time_stamp_postprocessed[-1][1]
+ })
+ return res
+ if len(punc_id_list) != len(time_stamp_postprocessed):
+ res.append({
+ 'text': text_postprocessed.split(),
+ "start": time_stamp_postprocessed[0][0],
+ "end": time_stamp_postprocessed[-1][1]
+ })
+ return res
+
+ sentence_text = ''
+ sentence_start = time_stamp_postprocessed[0][0]
+ texts = text_postprocessed.split()
+ for i in range(len(punc_id_list)):
+ sentence_text += texts[i]
+ if punc_id_list[i] == 2:
+ sentence_text += ','
+ res.append({
+ 'text': sentence_text,
+ "start": sentence_start,
+ "end": time_stamp_postprocessed[i][1]
+ })
+ sentence_text = ''
+ sentence_start = time_stamp_postprocessed[i][1]
+ elif punc_id_list[i] == 3:
+ sentence_text += '.'
+ res.append({
+ 'text': sentence_text,
+ "start": sentence_start,
+ "end": time_stamp_postprocessed[i][1]
+ })
+ sentence_text = ''
+ sentence_start = time_stamp_postprocessed[i][1]
+ return res
+
+
+
--
Gitblit v1.9.1