diff --git a/wenet/dataset/processor.py b/wenet/dataset/processor.py
index 8042c4a65..b69ceca85 100644
--- a/wenet/dataset/processor.py
+++ b/wenet/dataset/processor.py
@@ -25,6 +25,8 @@
 import torchaudio.compliance.kaldi as kaldi
 from torch.nn.utils.rnn import pad_sequence
 
+torchaudio.utils.sox_utils.set_buffer_size(16500)
+
 AUDIO_FORMAT_SETS = set(['flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma'])