A Cascaded Broadcast News Highlighter

Abstract

This paper presents a fully automatic news skimming system which takes a broadcast news audio stream and provides the user with the segmented, structured and highlighted transcript. This constitutes a system with three different, cascading stages: converting the audio stream to text using an automatic speech recogniser, segmenting into utterances and stories and finally determining which utterance should be highlighted using a saliency score. Each stage must operate on the erroneous output from the previous stage in the system; an effect which is naturally amplified as the data progresses through the processing stages. We present a large corpus of transcribed broadcast news data enabling us to investigate to which degree information worth highlighting survives this cascading of processes. Both extrinsic and intrinsic experimental results indicate that mistakes in the story boundary detection has a strong impact on the quality of highlights, whereas erroneous utterance boundaries cause only minor problems. Further, the difference in transcription quality does not affect the overall performance greatly.

Extractive speech summarization using evaluation metric-related training criteria

“…It has been shown that speech recognition errors are the dominating factor for the performance degradation of spoken document summarization when using recognition transcripts instead of manual transcripts, whereas erroneous sentence boundaries cause relatively minor problems (Christensen et al, 2008). A straightforward remedy, apart from the many approaches improving recognition accuracy, might be to develop more robust representations for spoken documents.…”

“…Notice that the positional feature is excluded in this study because it is not general enough and would highly depend on the epoches and genres of spoken documents (Christensen et al, 2008;Lin et al, 2010). …”

“…It should also be mentioned that speech summarization presents opportunities that do not exist for text summarization; for instance, information cues about prosody/acoustics and emotion/ speakers can help the determination of the importance and structure of spoken documents (Christensen, Gotoh, & Renals, 2008;McKeown, Hirschberg, Galley, & Maskey, 2005). We thus set the goal at selecting the most representative sentences (or the associated audio segments) based on the speech recognition (erroneous) transcripts, as well as a rich set of lexical and non-lexical features, to form the summary for a given spoken document.…”

A margin-based discriminative modeling approach for extractive speech summarization

“…Besides traditional unsupervised summarization methods [3][4][5][6][7][8][9], such as those based on document structural, linguistic or prosodic information, proximity or significance measures and relevance scores to identify salient sentences, machinelearning approaches with supervised training have drawn much attention and been applied with good success in a wide arrange of summarization tasks [3][4][5][6][7][8][9]. Specifically, the problem of speech summarization can be formulated as follows: Construct a ranking model (summarizer) that assigns a decision score (or a posterior probability) of being included in the summary to each sentence of a spoken document to be summarized.…”

Meta Comments for Summarizing Meeting Speech

“…There has been a variety of work concerned with the summarization of broadcast news [3,8,14,19], voicemail messages [11], lectures [9,21] and spontaneous conversations [18,22]. In this paper we are concerned with the summarization of multiparty meetings.…”