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
onnx (#1473)
zhifu gao15c4709beb4b
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex edcede5bf..a18224f46 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -494,11 +494,19 @@ class AutoModel:export_dir = export_utils.export_onnx(model=model,data_in=data_list,+ quantize=quantize,+ fallback_num=fallback_num,+ calib_num=calib_num,+ opset_version=opset_version,**kwargs)else:export_dir = export_utils.export_torchscripts(model=model,data_in=data_list,+ quantize=quantize,+ fallback_num=fallback_num,+ calib_num=calib_num,+ opset_version=opset_version,**kwargs)return export_dir\ No newline at end of filediff --git a/funasr/models/bicif_paraformer/model.py b/funasr/models/bicif_paraformer/model.pyindex b93f93a88..9849c8c22 100644--- a/funasr/models/bicif_paraformer/model.py+++ b/funasr/models/bicif_paraformer/model.py@@ -359,13 +359,10 @@ class BiCifParaformer(Paraformer):decoder_class = tables.decoder_classes.get(kwargs["decoder"] + "Export")self.decoder = decoder_class(self.decoder, onnx=is_onnx)- from funasr.utils.torch_function import MakePadMaskfrom funasr.utils.torch_function import sequence_mask++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)- if is_onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)self.forward = self.export_forwarddiff --git a/funasr/models/ct_transformer_streaming/encoder.py b/funasr/models/ct_transformer_streaming/encoder.pyindex badf5f6a5..bf0b8b277 100644--- a/funasr/models/ct_transformer_streaming/encoder.py+++ b/funasr/models/ct_transformer_streaming/encoder.py@@ -416,13 +416,10 @@ class SANMVadEncoderExport(torch.nn.Module):self.model = modelself._output_size = model._output_size- from funasr.utils.torch_function import MakePadMaskfrom funasr.utils.torch_function import sequence_mask-- if onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)+++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)from funasr.models.sanm.attention import MultiHeadedAttentionSANMExportdiff --git a/funasr/models/paraformer/decoder.py b/funasr/models/paraformer/decoder.pyindex 59c6e1d0e..7c370ba0c 100644--- a/funasr/models/paraformer/decoder.py+++ b/funasr/models/paraformer/decoder.py@@ -628,14 +628,12 @@ class ParaformerSANMDecoderExport(torch.nn.Module):):super().__init__()# self.embed = model.embed #Embedding(model.embed, max_seq_len)- from funasr.utils.torch_function import MakePadMask+from funasr.utils.torch_function import sequence_maskself.model = model- if onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)from funasr.models.sanm.attention import MultiHeadedAttentionSANMDecoderExportfrom funasr.models.sanm.attention import MultiHeadedAttentionCrossAttExport@@ -763,14 +761,12 @@ class ParaformerSANMDecoderOnlineExport(torch.nn.Module):super().__init__()# self.embed = model.embed #Embedding(model.embed, max_seq_len)self.model = model- from funasr.utils.torch_function import MakePadMask+from funasr.utils.torch_function import sequence_maskself.model = model- if onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)from funasr.models.sanm.attention import MultiHeadedAttentionSANMDecoderExportfrom funasr.models.sanm.attention import MultiHeadedAttentionCrossAttExport@@ -1036,14 +1032,12 @@ class ParaformerDecoderSANExport(torch.nn.Module):# self.embed = model.embed #Embedding(model.embed, max_seq_len)self.model = model- from funasr.utils.torch_function import MakePadMask+from funasr.utils.torch_function import sequence_maskself.model = model- if onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)from funasr.models.transformer.decoder import DecoderLayerExportdiff --git a/funasr/models/paraformer/model.py b/funasr/models/paraformer/model.pyindex 2e2a36e06..41a1bf73e 100644--- a/funasr/models/paraformer/model.py+++ b/funasr/models/paraformer/model.py@@ -566,15 +566,11 @@ class Paraformer(torch.nn.Module):decoder_class = tables.decoder_classes.get(kwargs["decoder"]+"Export")self.decoder = decoder_class(self.decoder, onnx=is_onnx)- from funasr.utils.torch_function import MakePadMaskfrom funasr.utils.torch_function import sequence_mask--- if is_onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)+self.forward = self.export_forwardreturn selfdiff --git a/funasr/models/paraformer_streaming/model.py b/funasr/models/paraformer_streaming/model.pyindex 518fe9369..33ec9762c 100644--- a/funasr/models/paraformer_streaming/model.py+++ b/funasr/models/paraformer_streaming/model.py@@ -579,14 +579,10 @@ class ParaformerStreaming(Paraformer):decoder_class = tables.decoder_classes.get(kwargs["decoder"] + "Export")self.decoder = decoder_class(self.decoder, onnx=is_onnx)- from funasr.utils.torch_function import MakePadMaskfrom funasr.utils.torch_function import sequence_mask-- if is_onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)-+++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)import copyimport typesdiff --git a/funasr/models/sanm/encoder.py b/funasr/models/sanm/encoder.pyindex f0a37227d..f5748187a 100644--- a/funasr/models/sanm/encoder.py+++ b/funasr/models/sanm/encoder.py@@ -503,13 +503,11 @@ class SANMEncoderExport(nn.Module):self.feats_dim = feats_dimself._output_size = model._output_size- from funasr.utils.torch_function import MakePadMask+from funasr.utils.torch_function import sequence_mask-- if onnx:- self.make_pad_mask = MakePadMask(max_seq_len, flip=False)- else:- self.make_pad_mask = sequence_mask(max_seq_len, flip=False)+++ self.make_pad_mask = sequence_mask(max_seq_len, flip=False)from funasr.models.sanm.attention import MultiHeadedAttentionSANMExportif hasattr(model, 'encoders0'):diff --git a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py b/runtime/python/onnxruntime/funasr_onnx/vad_bin.pyindex 384f3779e..6b3a1bcfe 100644--- a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py+++ b/runtime/python/onnxruntime/funasr_onnx/vad_bin.py@@ -63,8 +63,8 @@ class Fsmn_vad():model = AutoModel(model=model_dir)model_dir = model.export(type="onnx", quantize=quantize)- config_file = os.path.join(model_dir, 'vad.yaml')- cmvn_file = os.path.join(model_dir, 'vad.mvn')+ config_file = os.path.join(model_dir, 'config.yaml')+ cmvn_file = os.path.join(model_dir, 'am.mvn')config = read_yaml(config_file)self.frontend = WavFrontend(@@ -73,8 +73,8 @@ class Fsmn_vad():)self.ort_infer = OrtInferSession(model_file, device_id, intra_op_num_threads=intra_op_num_threads)self.batch_size = batch_size- self.vad_scorer = E2EVadModel(config["vad_post_conf"])- self.max_end_sil = max_end_sil if max_end_sil is not None else config["vad_post_conf"]["max_end_silence_time"]+ self.vad_scorer = E2EVadModel(config["model_conf"])+ self.max_end_sil = max_end_sil if max_end_sil is not None else config["model_conf"]["max_end_silence_time"]self.encoder_conf = config["encoder_conf"]def prepare_cache(self, in_cache: list = []):@@ -228,8 +228,8 @@ class Fsmn_vad_online():model = AutoModel(model=model_dir)model_dir = model.export(type="onnx", quantize=quantize)- config_file = os.path.join(model_dir, 'vad.yaml')- cmvn_file = os.path.join(model_dir, 'vad.mvn')+ config_file = os.path.join(model_dir, 'config.yaml')+ cmvn_file = os.path.join(model_dir, 'am.mvn')config = read_yaml(config_file)self.frontend = WavFrontendOnline(@@ -238,8 +238,8 @@ class Fsmn_vad_online():)self.ort_infer = OrtInferSession(model_file, device_id, intra_op_num_threads=intra_op_num_threads)self.batch_size = batch_size- self.vad_scorer = E2EVadModel(config["vad_post_conf"])- self.max_end_sil = max_end_sil if max_end_sil is not None else config["vad_post_conf"]["max_end_silence_time"]+ self.vad_scorer = E2EVadModel(config["model_conf"])+ self.max_end_sil = max_end_sil if max_end_sil is not None else config["model_conf"]["max_end_silence_time"]self.encoder_conf = config["encoder_conf"]def prepare_cache(self, in_cache: list = []):