Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
Update timestamp_tools.py
querrytondf768884a28b
diff --git a/funasr/utils/timestamp_tools.py b/funasr/utils/timestamp_tools.pyindex 32f0f8488..232137dba 100644--- a/funasr/utils/timestamp_tools.py+++ b/funasr/utils/timestamp_tools.py@@ -43,18 +43,18 @@ def ts_prediction_lfr6_standard(us_alphas,alphas, peaks = us_alphas, us_peaksif char_list[-1] == '</s>':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 offsetif len(fire_place) != len(char_list) + 1:alphas /= (alphas.sum() / (len(char_list) + 1))alphas = alphas.unsqueeze(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 offsetnum_frames = peaks.shape[0]timestamp_list = []new_char_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(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# begin silenceif fire_place[0] > START_END_THRESHOLD: