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9afcf0ea7d28
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex f5d6a82ab..2a55cd6bb 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -410,15 +410,17 @@ class LLMASR2(nn.Module):freeze_layer_num = int(audio_encoder_conf.get("freeze_layer_num", -1))if freeze_layer_num > 0:freeze_layer_num = range(freeze_layer_num)- else:- freeze_layer_num = [freeze_layer_num]+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 layer_id in freeze_layer_num:+ if isinstance(freeze_layer_num, (list, tuple)):+ if layer_id in freeze_layer_num:+ param.requires_grad = False+ else:param.requires_grad = Falseaudio_encoder.eval()