Exploring speaker characteristics for meeting summarization

Fei Liu, Yang Liu

Abstract

Summarizing spontaneous meetings is significantly more challenging than summarizing structured text like news articles due to disfluencies, lack of formal structure, and multi-party interactions. This paper investigates the hypothesis that individual speaker characteristics—the "who" behind the "what"—are critical for identifying salient information. By implementing speaker-sensitive feature normalization, the researchers demonstrate consistent improvements in extractive summarization performance across both human-transcribed and noisy automatic speech recognition (ASR) transcripts.

Introduction

Standard extractive summarization systems typically treat a meeting transcript as a monolithic block of text. However, meetings are inherently social and participant-driven. Every speaker has a unique style: some are verbose, others concise; some lead with their main points, while others build toward a conclusion. This study addresses the gap in "speaker-blind" summarization by investigating how individual patterns in sentence length, location, and similarity can be leveraged to better identify summary-worthy content.

Methods: Speaker-Sensitive Normalization

The authors frame summarization as a binary classification task for every Dialogue Act (DA). The core innovation is Speaker-Sensitive Feature Normalization. Rather than normalizing features against the entire meeting, the system compares them across three scales:

  1. Meeting Scale (mt): Features are normalized against the maximum value in the entire meeting.
  2. Speaker Scale (spkr): Features are normalized against the maximum value for that specific speaker, accounting for individual verbosity.
  3. Turn Scale (turn): Features are normalized within a single continuous speaking turn.

Statistical analysis using the Fisher Ratio confirmed that speaker-level normalization often provides a stronger signal for distinguishing summary sentences than meeting-level normalization, particularly for length and location features.

Fisher Ratio and Absolute Difference for various normalization scales across length, location, and similarity features.

Experimental Setup

The study utilized the ICSI Meeting Corpus, consisting of 75 naturally occurring academic meetings. The system was evaluated on:

  • Human Transcripts: High-quality manual transcriptions.
  • ASR Output: Noisy machine-generated text with a high Word Error Rate (38.2%).

A Maximum Entropy classifier was trained on 48 meetings, using features like normalized length, location (relative to the start and end of a speaker's total contribution), and cosine similarity to the meeting's centroid.

Key Insights

  • Relative Verbosity: A speaker's "verboseness" is a vital relative signal. A long sentence from a typically quiet person is often more significant than a long sentence from a consistently verbose participant.
  • Internal Turn Structure: Important points tend to cluster at the beginning or end of a speaker's total contribution. Normalizing location to the "speaker scale" captures this logic better than the meeting's overall timeline.
  • Filtering and Biographics: Filtering out "backchannels" (e.g., "uh-huh") removed 56% of non-summary text with minimal loss of content. While biographic attributes like gender or professional role showed slight statistical correlations, they did not significantly improve model performance beyond style-normalized features in this specific dataset.

Results

The "Speaker-norm" approach consistently outperformed the standard supervised baseline. On human transcripts, ROUGE-1 scores improved from 70.54 to 71.81, and the Pyramid score—a measure of content importance—jumped from 55.70 to 57.94. Notably, these improvements remained robust even on noisy ASR output, suggesting that speaker-level patterns are a resilient signal that survives poor speech-to-text quality.

Comprehensive performance comparison (ROUGE, Pyramid, and DA F-score) for human and ASR transcripts, showing the gains from speaker-sensitive normalization.

Conclusion

This research highlights that accounting for individual speaking styles is a powerful and robust method for improving meeting summarization. Future work may explore more complex modeling of speaker interactions and the impact of professional hierarchies in corporate environments, where biographic roles might play a more significant part in determining information saliency.

Summary updated

BibTeX
@inproceedings{liu2010exploring,
  author = {Fei Liu and Yang Liu},
  title = {{Exploring speaker characteristics for meeting summarization}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-676},
}
Figure