Leveraging Audio LLMs for Data Fault Detection in Audio Datasets

With the rapid development of artificial intelligence, model training with high-quality datasets is becoming increasingly important. Unfortunately, due to poor data quality or manual labeling errors, the dataset may contain data faults. Many methods have been proposed to detect data faults comparing the differences in sample features or loss functions existing between clean samples and fault samples. In this paper, we introduce a method for detecting data faults in audio datasets using the external knowledge of audio large language model. It contains three modules: designing three questions templates based on three different audio tasks, using audio LLMs to answer questions and evaluating the consistency of answers and labels. We evaluate the data fault detection performances on the audio datasets, including ESC-50, CREMAD and Free Spoken Digit Dataset and achieved a high TPR and a low FPR. Our method achieves state-of-the-art performance and performs well in different audio classification tasks.