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
游雁ec42889511b1
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.pyindex 3b70ad6d4..48a983c16 100644--- a/funasr/auto/auto_model.py+++ b/funasr/auto/auto_model.py@@ -181,15 +181,18 @@ class AutoModel:# init_paraminit_param = kwargs.get("init_param", None)if init_param is not None:- logging.info(f"Loading pretrained params from {init_param}")- load_pretrained_model(- model=model,- path=init_param,- ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),- oss_bucket=kwargs.get("oss_bucket", None),- scope_map=kwargs.get("scope_map", None),- excludes=kwargs.get("excludes", None),- )+ if os.path.exists(init_param):+ logging.info(f"Loading pretrained params from {init_param}")+ load_pretrained_model(+ model=model,+ path=init_param,+ ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),+ oss_bucket=kwargs.get("oss_bucket", None),+ scope_map=kwargs.get("scope_map", None),+ excludes=kwargs.get("excludes", None),+ )+ else:+ print(f"error, init_param does not exist!: {init_param}")return model, kwargsdiff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex e3cf55173..937c8e1bf 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -217,7 +217,7 @@ class LLMASRNAR(nn.Module):) -> Tuple[torch.Tensor, torch.Tensor]:audio_mask = kwargs.get("audio_mask")- audio_token_lengths = audio_mask.sum(-1)+ audio_token_lengths = audio_mask.sum(-1) if audio_mask else Nonebatch = {"speech": speech, "speech_lengths": speech_lengths}enc, enc_lens = self.audio_encoder.encode(**batch)@@ -279,7 +279,7 @@ class LLMASRNAR(nn.Module):prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(prompt)- prompt_ids = self.tokenizer.encode(prompt_pre)+ prompt_ids = tokenizer.encode(prompt_pre)prompt_length = len(prompt_ids)prompt_ids = torch.tensor(prompt_ids, dtype=torch.int64).to(kwargs["device"])diff --git a/funasr/train_utils/load_pretrained_model.py b/funasr/train_utils/load_pretrained_model.pyindex 8493bf58c..03a6ff5ed 100644--- a/funasr/train_utils/load_pretrained_model.py+++ b/funasr/train_utils/load_pretrained_model.py@@ -118,7 +118,7 @@ def load_pretrained_model(else:print(f"Warning, miss key in ckpt: {k}, mapped: {k_ddp}")- flag = obj.load_state_dict(dst_state, strict=True)+ flag = obj.load_state_dict(dst_state, strict=False)# print(flag)# def load_pretrained_model(