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

Update timestamp_tools.py

Browse Source 
[fix] fix a bug in function "ts_prediction_lfr6_standard"
This commit is contained in:
querryton 2024-04-19 13:37:05 +08:00 committed by GitHub
parent 8b03379434
commit df768884a2
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
funasr/utils/timestamp_tools.py
View File

@ -43,18 +43,18 @@ def ts_prediction_lfr6_standard(us_alphas,
alphas, peaks = us_alphas, us_peaks alphas, peaks = us_alphas, us_peaks
if char_list[-1] == '</s>': if char_list[-1] == '</s>':
char_list = char_list[:-1] char_list = char_list[:-1]
fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift # total offset fire_place = torch.where(peaks>=1.0-1e-4)[0].cpu().numpy() + force_time_shift # total offset
if len(fire_place) != len(char_list) + 1: if len(fire_place) != len(char_list) + 1:
alphas /= (alphas.sum() / (len(char_list) + 1)) alphas /= (alphas.sum() / (len(char_list) + 1))
alphas = alphas.unsqueeze(0) alphas = alphas.unsqueeze(0)
peaks = cif_wo_hidden(alphas, threshold=1.0-1e-4)[0] peaks = cif_wo_hidden(alphas, threshold=1.0-1e-4)[0]
fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift # total offset fire_place = torch.where(peaks>=1.0-1e-4)[0].cpu().numpy() + force_time_shift # total offset
num_frames = peaks.shape[0] num_frames = peaks.shape[0]
timestamp_list = [] timestamp_list = []
new_char_list = [] new_char_list = []
# for bicif model trained with large data, cif2 actually fires when a character starts # 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 # so treat the frames between two peaks as the duration of the former token
fire_place = torch.where(peaks>1.0-1e-4)[0].cpu().numpy() + force_time_shift # total offset # fire_place = torch.where(peaks>=1.0-1e-4)[0].cpu().numpy() + force_time_shift # total offset
# assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1 # assert num_peak == len(char_list) + 1 # number of peaks is supposed to be number of tokens + 1
# begin silence # begin silence
if fire_place[0] > START_END_THRESHOLD: if fire_place[0] > START_END_THRESHOLD: