ASR-Based Audio Pattern Discovery

we present an ASR-based approach based to detect the audio pattern (repeating words or multiword phrases etc). We take full advantage of recognition results produced by our Chinese ASR system and compare the intermediate results of audios to generate word patterns. We align the intermediate results with S-DTW algorithm to retrieve the repeating audio pattern in the audio stream. And then, the graph clustering algorithm will be used to cluster the word patterns together. To achieve the higher purity, acoustic verification is applied to rescore cluster results. We evaluate our system´s performance on both BN speech and CTS speech. On BN speech, we achieve the word pattern purity of 94.73% and on CTS speech we achieve the word pattern purity of 91.66%. We also analyze the relationship between the recognition performance and pattern clustering performance.