Abstract
Automatic speech summarization faces unique challenges compared to text summarization, including the lack of explicit formatting and the presence of acoustic noise and disfluencies. This paper investigates the effectiveness of various feature sets—acoustic, structural, and lexical—across two distinct speech genres: Mandarin Broadcast News (BN) and Lecture Speech (LS). Using a Support Vector Machine (SVM) framework, the study demonstrates that while structural cues are paramount for the highly formatted broadcast news, lexical content is more critical for spontaneous lecture speech. Furthermore, the research reveals a significant resilience of summarization performance to errors in Automatic Speech Recognition (ASR), suggesting that perfect transcription is not a prerequisite for high-quality extractive summaries.
Introduction
Summarizing spoken content is inherently more complex than summarizing written documents. Speech lacks the structural markers found in text, such as paragraphs and headings, and is often characterized by spontaneous disfluencies like "ums" and "ahs." Most prior research has focused on single-genre summarization. This study aims to bridge this gap by comparing Mandarin Broadcast News, which is professionally produced and highly structured, with Lecture Speech, which is more spontaneous and content-driven. The goal is to identify universal versus genre-specific features for extractive summarization.
Methodology
The authors frame extractive summarization as a binary classification task, where an SVM with a Radial Basis Function (RBF) kernel determines whether a given sentence should be included in the summary. The classification relies on three primary feature categories:
- Acoustic/Prosodic Features: These capture the vocal characteristics of the speaker. The system utilizes 13 specific features, including pitch (F0) statistics, energy levels, and speaking rates.
- Structural Features: These leverage the inherent organization of the speech. For BN, this includes sentence position and speaker turns. A "Poisson Noun" feature is also employed to identify key topical nouns based on their statistical distribution.
- Lexical Features: These are traditional text-based metrics, such as TF-IDF scores, sentence length, and the frequency of Named Entities.
Experimental Setup
The evaluation was conducted on two Mandarin Chinese datasets. The Broadcast News (BN) set consisted of 347 stories from the 1997 Hub4 Mandarin corpus, while the Lecture Speech (LS) set included 60 presentations with manual transcripts. To assess real-world robustness, an in-house ASR system with a 21.8% Character Error Rate (CER) was used to generate transcripts for comparison against manual versions.
Results and Discussion
The performance was measured using the ROUGE-L F-measure. The results highlight a clear divergence in optimal strategies for different genres. In Broadcast News, structural features proved to be the most potent, with the model achieving an F-measure of 0.64. For Lecture Speech, lexical features were more influential, yielding an F-measure of 0.65.
A key finding of this study is the "ASR-resilience" of the summarization process. As shown in Table 3, for lecture speech, the F-measure only decreased slightly from 0.623 (using manual transcripts, Le(M)) to 0.616 (using ASR transcripts, Le(A)). Analysis suggests that most ASR errors occur on single characters that do not significantly alter the semantic keywords necessary for identifying important sentences. Professional news anchors use specific pitch and energy patterns to highlight key information, making acoustic features much more useful for news than for the more varied speaking styles found in lectures.
Conclusion
This comparative study underscores that speech summarization strategies must be tailored to the genre. Professional speech like broadcast news benefits significantly from prosodic and structural cues, while spontaneous speech like lectures relies more on lexical content. Importantly, the robustness of these models to ASR errors indicates that automated summarization can be effective even with imperfect transcription technology, paving the way for practical applications in large-scale audio repositories.
Summary updated