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
test
语帆2bffe1d5392b
diff --git a/examples/industrial_data_pretraining/lcbnet/demo2.sh b/examples/industrial_data_pretraining/lcbnet/demo2.shindex cfb5b235e..0d5a4f031 100755--- a/examples/industrial_data_pretraining/lcbnet/demo2.sh+++ b/examples/industrial_data_pretraining/lcbnet/demo2.sh@@ -6,7 +6,6 @@ python -m funasr.bin.inference \--config-name="config.yaml" \++init_param=${file_dir}/model.pb \++tokenizer_conf.token_list=${file_dir}/tokens.txt \-++frontend_conf.cmvn_file=${file_dir}/am.mvn \++input=[${file_dir}/wav.scp,${file_dir}/ocr_text] \+data_type='["kaldi_ark", "text"]' \++tokenizer_conf.bpemodel=${file_dir}/bpe.model \diff --git a/funasr/models/lcbnet/model.py b/funasr/models/lcbnet/model.pyindex 1acce785f..f45e71d6f 100644--- a/funasr/models/lcbnet/model.py+++ b/funasr/models/lcbnet/model.py@@ -21,6 +21,7 @@ from funasr.utils.load_utils import load_audio_text_image_video, extract_fbankfrom funasr.utils import postprocess_utilsfrom funasr.utils.datadir_writer import DatadirWriterfrom funasr.register import tables+import pdb@tables.register("model_classes", "LCBNet")class LCBNet(nn.Module):@@ -92,6 +93,7 @@ class LCBNet(nn.Module):bias_predictor_class = tables.encoder_classes.get(bias_predictor)bias_predictor = bias_predictor_class(**bias_predictor_conf)+if decoder is not None:decoder_class = tables.decoder_classes.get(decoder)decoder = decoder_class(@@ -272,15 +274,15 @@ class LCBNet(nn.Module):ind: int"""with autocast(False):-+ pdb.set_trace()# Data augmentationif self.specaug is not None and self.training:speech, speech_lengths = self.specaug(speech, speech_lengths)-+ pdb.set_trace()# Normalization for feature: e.g. Global-CMVN, Utterance-CMVNif self.normalize is not None:speech, speech_lengths = self.normalize(speech, speech_lengths)-+ pdb.set_trace()# Forward encoder# feats: (Batch, Length, Dim)# -> encoder_out: (Batch, Length2, Dim2)@@ -297,7 +299,7 @@ class LCBNet(nn.Module):if intermediate_outs is not None:return (encoder_out, intermediate_outs), encoder_out_lens-+ pdb.set_trace()return encoder_out, encoder_out_lensdef _calc_att_loss(@@ -442,6 +444,7 @@ class LCBNet(nn.Module):speech = speech.to(device=kwargs["device"])speech_lengths = speech_lengths.to(device=kwargs["device"])+ pdb.set_trace()# Encoderencoder_out, encoder_out_lens = self.encode(speech, speech_lengths)if isinstance(encoder_out, tuple):diff --git a/funasr/utils/load_utils.py b/funasr/utils/load_utils.pyindex 963f5c258..644af2324 100644--- a/funasr/utils/load_utils.py+++ b/funasr/utils/load_utils.py@@ -108,10 +108,7 @@ def extract_fbank(data, data_len = None, data_type: str="sound", frontend=None,data_list.append(data_i)data_len.append(data_i.shape[0])data = pad_sequence(data_list, batch_first=True) # data: [batch, N]- # import pdb;- # pdb.set_trace()- # if data_type == "sound":- pdb.set_trace()+data, data_len = frontend(data, data_len, **kwargs)if isinstance(data_len, (list, tuple)):