Diagnostic techniques for spoken keyword discovery

Keyword discovery is an unsupervised technology that can help to process collections of speech and capture repeated patterns. This technology becomes useful and provides solution for unsupervised content analysis tasks, especially when the acoustic and lexical characteristics are not known in advance or there is little or no data to model these characteristics via statistical models. In these situations, keyword discovery can find potentially important words for further analysis using minimal resources. Unfortunately, keyword discovery performance heavily depends on the quality of the features used to characterize the raw signal and the alignment algorithm used to find similar feature subsequences. It is not yet fully understood which features and alignment algorithms work well in different scenarios and for different tasks, and there are very few diagnostic techniques for improving our understanding. In this paper, we present two diagnostic measurements that can be used to directly assess the quality of alignments between sequences of features independently of the intended use of the alignments downstream. We argue that such diagnostic techniques are valuable for intrinsically assessing speech features and alignment algorithms for keyword detection.