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7d57828086d8
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex dd806cf66..e7dd11ea3 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -407,9 +407,19 @@ class LLMASR2(nn.Module):audio_encoder = encoder_class(input_size=input_size, **audio_encoder_conf)audio_encoder_output_size = audio_encoder.output_size()freeze = audio_encoder_conf.get("freeze", True)+ freeze_layer_num = 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():- param.requires_grad = False+ 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 = Falseaudio_encoder.eval()self.audio_encoder = audio_encoder