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
decoding
游雁e4a69d476867
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex fb0bee3ca..14837b936 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -413,15 +413,16 @@ class LLMASR2(nn.Module):if freeze:for name, param in audio_encoder.named_parameters():- idx = re.search(r"\.\d+\.", name)- if idx is not None:- beg, end = idx.regs[0]- layer_id = int(name[beg + 1 : end - 1])- if isinstance(freeze_layer_num, (list, tuple)):+ if isinstance(freeze_layer_num, (list, tuple)):+ idx = re.search(r"\.\d+\.", name)+ if idx is not None:+ beg, end = idx.regs[0]+ layer_id = int(name[beg + 1 : end - 1])if layer_id in freeze_layer_num:param.requires_grad = False- else:- param.requires_grad = False+ else:+ param.requires_grad = False+audio_encoder.eval()self.audio_encoder = audio_encoderdiff --git a/funasr/train_utils/trainer_ds.py b/funasr/train_utils/trainer_ds.pyindex b2d7b95d8..22be426e9 100644--- a/funasr/train_utils/trainer_ds.py+++ b/funasr/train_utils/trainer_ds.py@@ -313,6 +313,7 @@ class Trainer:state_dict = model.state_dict()if self.effective_save_name_excludes is not None:+ logging.info(f"effective_save_name_excludes: {self.effective_save_name_excludes}")dst_state_dict = {}for k in state_dict.keys():for k_ex in self.effective_save_name_excludes: