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
Dev gzf (#1467)
zhifu gaoa7d7a0f3a2e7
diff --git a/README.md b/README.mdindex 9841fe1e1..d15905025 100644--- a/README.md+++ b/README.md@@ -215,14 +215,14 @@ More examples ref to [docs](https://github.com/alibaba-damo-academy/FunASR/tree/### Command-line usage```shell-funasr-export ++model=paraformer ++quantize=false+funasr-export ++model=paraformer ++quantize=false ++device=cpu```-### python+### Python```pythonfrom funasr import AutoModel-model = AutoModel(model="paraformer")+model = AutoModel(model="paraformer", device="cpu")res = model.export(quantize=False)```diff --git a/examples/industrial_data_pretraining/bicif_paraformer/export.py b/examples/industrial_data_pretraining/bicif_paraformer/export.pyindex 138f23a69..c819f7a28 100644--- a/examples/industrial_data_pretraining/bicif_paraformer/export.py+++ b/examples/industrial_data_pretraining/bicif_paraformer/export.py@@ -8,7 +8,7 @@from funasr import AutoModelmodel = AutoModel(model="damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch",- model_revision="v2.0.4")+ model_revision="v2.0.4", device="cpu")res = model.export(type="onnx", quantize=False)print(res)@@ -17,7 +17,7 @@ print(res)# method2, inference from local pathfrom funasr import AutoModel-model = AutoModel(model="/Users/zhifu/.cache/modelscope/hub/iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")+model = AutoModel(model="/Users/zhifu/.cache/modelscope/hub/iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch", device="cpu")res = model.export(type="onnx", quantize=False)print(res)\ No newline at end of filediff --git a/examples/industrial_data_pretraining/bicif_paraformer/export.sh b/examples/industrial_data_pretraining/bicif_paraformer/export.shindex 42b63484b..b6883b7c9 100644--- a/examples/industrial_data_pretraining/bicif_paraformer/export.sh+++ b/examples/industrial_data_pretraining/bicif_paraformer/export.sh@@ -12,7 +12,8 @@ python -m funasr.bin.export \++model=${model} \++model_revision=${model_revision} \++type="onnx" \-++quantize=false+++quantize=false \+++device="cpu"# method2, inference from local pathmodel="/Users/zhifu/.cache/modelscope/hub/iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"@@ -20,4 +21,5 @@ model="/Users/zhifu/.cache/modelscope/hub/iic/speech_paraformer-large_asr_nat-zhpython -m funasr.bin.export \++model=${model} \++type="onnx" \-++quantize=false\ No newline at end of file+++quantize=false \+++device="cpu"\ No newline at end of filediff --git a/examples/industrial_data_pretraining/ct_transformer/export.sh b/examples/industrial_data_pretraining/ct_transformer/export.shindex 755645855..a11cda5d4 100644--- a/examples/industrial_data_pretraining/ct_transformer/export.sh+++ b/examples/industrial_data_pretraining/ct_transformer/export.sh@@ -12,7 +12,8 @@ python -m funasr.bin.export \++model=${model} \++model_revision=${model_revision} \++type="onnx" \-++quantize=false+++quantize=false \+++device="cpu"# method2, inference from local path@@ -21,4 +22,5 @@ model="/Users/zhifu/.cache/modelscope/hub/iic/punc_ct-transformer_zh-cn-common-vpython -m funasr.bin.export \++model=${model} \++type="onnx" \-++quantize=false\ No newline at end of file+++quantize=false \+++device="cpu"\ No newline at end of filediff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex c4bab03e2..28b9e94fc 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -156,7 +156,7 @@ class AutoModel:kwargs["batch_size"] = 1kwargs["device"] = device- if kwargs.get("ncpu", None):+ if kwargs.get("ncpu", 4):torch.set_num_threads(kwargs.get("ncpu"))# build tokenizer@@ -476,11 +476,13 @@ class AutoModel:calib_num: int = 100,opset_version: int = 14,**cfg):- os.environ['EXPORTING_MODEL'] = 'TRUE'++ device = cfg.get("device", "cpu")+ model = self.model.to(device=device)kwargs = self.kwargsdeep_update(kwargs, cfg)+ kwargs["device"] = devicedel kwargs["model"]- model = self.modelmodel.eval()batch_size = 1diff --git a/funasr/bin/export.py b/funasr/bin/export.pyindex 7d476644a..cb160e9c5 100644--- a/funasr/bin/export.py+++ b/funasr/bin/export.py@@ -24,7 +24,8 @@ def main_hydra(cfg: DictConfig):if kwargs.get("debug", False):import pdb; pdb.set_trace()-+ if "device" not in kwargs:+ kwargs["device"] = "cpu"model = AutoModel(**kwargs)res = model.export(input=kwargs.get("input", None),diff --git a/funasr/models/bicif_paraformer/model.py b/funasr/models/bicif_paraformer/model.pyindex eb7318b20..4802da07d 100644--- a/funasr/models/bicif_paraformer/model.py+++ b/funasr/models/bicif_paraformer/model.py@@ -23,7 +23,7 @@ from funasr.models.transformer.utils.add_sos_eos import add_sos_eosfrom funasr.utils.timestamp_tools import ts_prediction_lfr6_standardfrom funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_listfrom funasr.utils.load_utils import load_audio_text_image_video, extract_fbank-+from funasr.train_utils.device_funcs import to_deviceif LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):from torch.cuda.amp import autocast@@ -348,6 +348,7 @@ class BiCifParaformer(Paraformer):max_seq_len=512,**kwargs,):+ self.device = kwargs.get("device")is_onnx = kwargs.get("type", "onnx") == "onnx"encoder_class = tables.encoder_classes.get(kwargs["encoder"] + "Export")self.encoder = encoder_class(self.encoder, onnx=is_onnx)@@ -370,14 +371,14 @@ class BiCifParaformer(Paraformer):return self- def _export_forward(+ def export_forward(self,speech: torch.Tensor,speech_lengths: torch.Tensor,):# a. To devicebatch = {"speech": speech, "speech_lengths": speech_lengths}- # batch = to_device(batch, device=self.device)+ batch = to_device(batch, device=self.device)enc, enc_len = self.encoder(**batch)mask = self.make_pad_mask(enc_len)[:, None, :]diff --git a/funasr/models/ct_transformer/model.py b/funasr/models/ct_transformer/model.pyindex 31b8c27d5..88ee86734 100644--- a/funasr/models/ct_transformer/model.py+++ b/funasr/models/ct_transformer/model.py@@ -18,7 +18,6 @@ from funasr.utils.load_utils import load_audio_text_image_videofrom funasr.models.transformer.utils.nets_utils import make_pad_maskfrom funasr.models.ct_transformer.utils import split_to_mini_sentence, split_words-if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):from torch.cuda.amp import autocastelse:@@ -378,7 +377,7 @@ class CTTransformer(torch.nn.Module):return self- def _export_forward(self, inputs: torch.Tensor, text_lengths: torch.Tensor):+ def export_forward(self, inputs: torch.Tensor, text_lengths: torch.Tensor):"""Compute loss value from buffer sequences.Args:diff --git a/funasr/models/ct_transformer_streaming/model.py b/funasr/models/ct_transformer_streaming/model.pyindex a9b2efb82..4752c4b5a 100644--- a/funasr/models/ct_transformer_streaming/model.py+++ b/funasr/models/ct_transformer_streaming/model.py@@ -186,7 +186,7 @@ class CTTransformerStreaming(CTTransformer):return self- def _export_forward(self, inputs: torch.Tensor,+ def export_forward(self, inputs: torch.Tensor,text_lengths: torch.Tensor,vad_indexes: torch.Tensor,sub_masks: torch.Tensor,diff --git a/funasr/models/fsmn_vad_streaming/model.py b/funasr/models/fsmn_vad_streaming/model.pyindex c3063b0a8..d06db2093 100644--- a/funasr/models/fsmn_vad_streaming/model.py+++ b/funasr/models/fsmn_vad_streaming/model.py@@ -651,7 +651,7 @@ class FsmnVADStreaming(nn.Module):return self- def _export_forward(self, feats: torch.Tensor, *args, **kwargs):+ def export_forward(self, feats: torch.Tensor, *args, **kwargs):scores, out_caches = self.encoder(feats, *args)diff --git a/funasr/models/paraformer/model.py b/funasr/models/paraformer/model.pyindex 586d72de0..f5f0e4e22 100644--- a/funasr/models/paraformer/model.py+++ b/funasr/models/paraformer/model.py@@ -21,7 +21,7 @@ from funasr.losses.label_smoothing_loss import LabelSmoothingLossfrom funasr.models.transformer.utils.add_sos_eos import add_sos_eosfrom funasr.models.transformer.utils.nets_utils import make_pad_mask, pad_listfrom funasr.utils.load_utils import load_audio_text_image_video, extract_fbank-+from funasr.train_utils.device_funcs import to_device@tables.register("model_classes", "Paraformer")class Paraformer(torch.nn.Module):@@ -554,7 +554,7 @@ class Paraformer(torch.nn.Module):max_seq_len=512,**kwargs,):-+ self.device = kwargs.get("device")is_onnx = kwargs.get("type", "onnx") == "onnx"encoder_class = tables.encoder_classes.get(kwargs["encoder"]+"Export")self.encoder = encoder_class(self.encoder, onnx=is_onnx)@@ -579,14 +579,14 @@ class Paraformer(torch.nn.Module):return self- def _export_forward(+ def export_forward(self,speech: torch.Tensor,speech_lengths: torch.Tensor,):# a. To devicebatch = {"speech": speech, "speech_lengths": speech_lengths}- # batch = to_device(batch, device=self.device)+ batch = to_device(batch, device=self.device)enc, enc_len = self.encoder(**batch)mask = self.make_pad_mask(enc_len)[:, None, :]diff --git a/funasr/models/paraformer_streaming/model.py b/funasr/models/paraformer_streaming/model.pyindex cebbfc16d..63dba5d88 100644--- a/funasr/models/paraformer_streaming/model.py+++ b/funasr/models/paraformer_streaming/model.py@@ -566,7 +566,7 @@ class ParaformerStreaming(Paraformer):max_seq_len=512,**kwargs,):-+ self.device = kwargs.get("device")is_onnx = kwargs.get("type", "onnx") == "onnx"encoder_class = tables.encoder_classes.get(kwargs["encoder"] + "Export")self.encoder = encoder_class(self.encoder, onnx=is_onnx)@@ -612,7 +612,7 @@ class ParaformerStreaming(Paraformer):return encoder_model, decoder_model- def _export_encoder_forward(+ def export_encoder_forward(self,speech: torch.Tensor,speech_lengths: torch.Tensor,@@ -663,7 +663,7 @@ class ParaformerStreaming(Paraformer):def export_encoder_name(self):return "model.onnx"- def _export_decoder_forward(+ def export_decoder_forward(self,enc: torch.Tensor,enc_len: torch.Tensor,diff --git a/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py b/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.pyindex e047db936..82548ad9c 100644--- a/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py+++ b/runtime/python/onnxruntime/funasr_onnx/paraformer_bin.py@@ -63,8 +63,8 @@ class Paraformer():"For the users in China, you could install with the command:\n" \"\npip3 install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple"- model = AutoModel(model=cache_dir)- model_dir = model.export(type="onnx", quantize=quantize, device="cpu")+ model = AutoModel(model=model_dir)+ model_dir = model.export(type="onnx", quantize=quantize)config_file = os.path.join(model_dir, 'config.yaml')cmvn_file = os.path.join(model_dir, 'am.mvn')diff --git a/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py b/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.pyindex 7da5afc6a..692596079 100644--- a/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py+++ b/runtime/python/onnxruntime/funasr_onnx/paraformer_online_bin.py@@ -55,8 +55,8 @@ class Paraformer():"For the users in China, you could install with the command:\n" \"\npip3 install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple"- model = AutoModel(model=cache_dir)- model_dir = model.export(type="onnx", quantize=quantize, device="cpu")+ model = AutoModel(model=model_dir)+ model_dir = model.export(type="onnx", quantize=quantize)config_file = os.path.join(model_dir, 'config.yaml')cmvn_file = os.path.join(model_dir, 'am.mvn')diff --git a/runtime/python/onnxruntime/funasr_onnx/punc_bin.py b/runtime/python/onnxruntime/funasr_onnx/punc_bin.pyindex 4e1014f07..b1aca6ea2 100644--- a/runtime/python/onnxruntime/funasr_onnx/punc_bin.py+++ b/runtime/python/onnxruntime/funasr_onnx/punc_bin.py@@ -55,8 +55,8 @@ class CT_Transformer():"For the users in China, you could install with the command:\n" \"\npip3 install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple"- model = AutoModel(model=cache_dir)- model_dir = model.export(type="onnx", quantize=quantize, device="cpu")+ model = AutoModel(model=model_dir)+ model_dir = model.export(type="onnx", quantize=quantize)config_file = os.path.join(model_dir, 'punc.yaml')config = read_yaml(config_file)diff --git a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py b/runtime/python/onnxruntime/funasr_onnx/vad_bin.pyindex af32b1d53..384f3779e 100644--- a/runtime/python/onnxruntime/funasr_onnx/vad_bin.py+++ b/runtime/python/onnxruntime/funasr_onnx/vad_bin.py@@ -61,8 +61,8 @@ class Fsmn_vad():"For the users in China, you could install with the command:\n" \"\npip3 install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple"- model = AutoModel(model=cache_dir)- model_dir = model.export(type="onnx", quantize=quantize, device="cpu")+ 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 = read_yaml(config_file)@@ -225,8 +225,8 @@ class Fsmn_vad_online():"For the users in China, you could install with the command:\n" \"\npip3 install -U funasr -i https://mirror.sjtu.edu.cn/pypi/web/simple"- model = AutoModel(model=cache_dir)- model_dir = model.export(type="onnx", quantize=quantize, device="cpu")+ 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')