Speech summarization using clustered frames acoustic units for Persian news

In this paper we present a method for news speech summarization that can select directly the important sentences from a spoken document. In this paper instead of using the conventional speech units such as phonemes and words, a set of new speech units obtained by a clustering technique are used. Our speech summarization method uses these new acoustic units instead of words for speech summarization. To obtain the new speech units, simply we extract frames from training set of summarization corpus and cluster them using K-means algorithm. To summarize a spoken document, each sentence of the document is represented by a sequences of new acoustic units. Then we use n-grams of new acoustic units for summarization. Experiments were conducted on a Persian news dataset, and encouraging results were obtained. The results were compared to speech summarization method using a phoneme recognition system according to ROUGE-N and ROUGE-L measures.