We propose a method to automatically discover repeating acoustic patterns in speech signals in an unsupervised manner, allowing variability between occurences of a pattern. The resulting patterns, known as audio motifs, are mostly words or sequences of words characterics of the audio content. In this paper, we formalize the problem of motif discovery in speech signals and describe a practical solution using DTW and exploiting the local repetitiveness of motifs. Experimental results on the motif discovery task are provided on a large radio broadcast news corpus. We also propose a refinement of the DTW-based method to account for more variability.