In this paper, we propose an adaptive audio watermarking scheme according to local audio features. Firstly, the original audio signal is partitioned into audio frames and these audio frames are transformed into DWT domain respectively. Next, the local features of each audio frame are extracted respectively, and these features are used to train kernel fuzzy c-means (KFCM) clustering algorithm. According to well-trained KFCM, the audio frames to embed the watermark are selected and their embedding strengths are determined adaptively. The experimental results show the proposed method is robust to common signal processing operations such as lossy compression (MP3), filtering, re-sampling, re-quantizing, etc.