Summarizing multiple spoken documents: finding evidence from untranscribed audio

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 549–557, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Summarizing multiple spoken documents: finding evidence from untranscribed audio Xiaodan Zhu, Gerald Penn and Frank Rudzicz University of Toronto 10 King’s College Rd., Toronto, M5S 3G4, ON, Canada {xzhu,gpenn,frank}@cs.toronto.edu Abstract This paper presents a model for summa- rizing multiple untranscribed spoken doc- uments. Without assuming the availabil- ity of transcripts, the model modifies a recently proposed unsupervised algorithm to detect re-occurring acoustic patterns in speech and uses them to estimate similari- ties between utterances, which are in turn used to identify salient utterances and re- move redundancies. This model is of in- terest due to its independence from spo- ken language transcription, an error-prone and resource-intensive process, its abil- ity to integrate multiple sources of infor- mation on the same topic, and its novel use of acoustic patterns that extends pre- vious work on low-level prosodic feature detection. We compare the performance of this model with that achieved using man- ual and automatic transcripts, and find that this new approach is roughly equivalent to having access to ASR transcripts with word error rates in the 33–37% range with- out actually having to do the ASR, plus it better handles utterances with out-of- vocabulary words. 1 Introduction Summarizing spoken documents has been exten- sively studied over the past several years (Penn and Zhu, 2008; Maskey and Hirschberg, 2005; Murray et al., 2005; Christensen et al., 2004; Zechner, 2001). Conventionally called speech summarization, although speech connotes more than spoken documents themselves, it is motivated by the demand for better ways to navigate spoken content and the natural difficulty in doing so — speech is inherently more linear or sequential than text in its traditional delivery. Previous research on speech summarization has addressed several important problems in this field (see Section 2.1). All of this work, however, has focused on single-document summarization and the integration of fairly simplistic acoustic features, inspired by work in descriptive linguis- tics. The issues of navigating speech content are magnified when dealing with larger collections — multiple spoken documents on the same topic. For example, when one is browsing news broadcasts covering the same events or call-centre record- ings related to the same type of customer ques- tions, content redundancy is a prominent issue. Multi-document summarization on written docu- ments has been studied for more than a decade (see Section 2.2). Unfortunately, no such effort has been made on audio documents yet. An obvious way to summarize multiple spo- ken documents is to adopt the transcribe-and- summarize approach, in which automatic speech recognition (ASR) is first employed to acquire written transcripts. Speech summarization is ac- cordingly reduced to a text summarization task conducted on error-prone transcripts. Such an approach, however, encounters several problems. First, assuming the availability of ASR is not always valid for many languages other than English that one may want to summarize. Even when it is, transcription quality …

Content truncated to 500 words