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7355e2050321
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex 6e7939b4c..80c33fc0c 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -496,11 +496,14 @@ class LLMASR2(nn.Module):batch_size, frames, _ = speech.shape- # audio encoder- encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)+ with torch.cuda.amp.autocast(enabled=False):+ # audio encoder+ encoder_out, encoder_out_lens = self.audio_encoder(+ speech.permute(0, 2, 1), speech_lengths+ )- # audio_adaptor- encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)+ # audio_adaptor+ encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)input_ids[input_ids < 0] = 0inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)