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
fix bug
游雁9fd3c47fd7ff
diff --git a/funasr/models/llm_asr/model.py b/funasr/models/llm_asr/model.pyindex f8c3efc77..18791d004 100644--- a/funasr/models/llm_asr/model.py+++ b/funasr/models/llm_asr/model.py@@ -19,6 +19,7 @@ from funasr.utils import postprocess_utilsfrom funasr.utils.datadir_writer import DatadirWriterfrom funasr.register import tablesfrom funasr.train_utils.device_funcs import to_device+import traceback@tables.register("model_classes", "LLMASR")@@ -489,6 +490,7 @@ class LLMASR2(nn.Module):fbank_fake_len = fbank_fake_lens[batch_idx].item()fbank_beg_idx = fbank_beg[batch_idx, 0].item()min_len = min(fbank_fake_len, inputs_embeds.shape[1] - fbank_beg_idx)+try:inputs_embeds[batch_idx, fbank_beg_idx : fbank_beg_idx + min_len, :] = encoder_out[batch_idx, :min_len, :@@ -499,7 +501,7 @@ class LLMASR2(nn.Module):f"batch_idx: {batch_idx}, inputs_embeds: {inputs_embeds.shape}, fbank_beg_idx: {fbank_beg_idx}, min_len: {min_len}, fbank_fake_len: {fbank_fake_len}")fbank_fake_len = encoder_out_lens[batch_idx].item()- min_len = min(fbank_fake_len, inputs_embeds.shape[1] - fbank_beg_idx)+ min_len = min(fbank_fake_len, min_len)inputs_embeds[batch_idx, fbank_beg_idx : fbank_beg_idx + min_len, :] = encoder_out[batch_idx, :min_len, :]