Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Paragraph Segmentation Revisited: Towards a Standard Task for Structuring Speech

Fabian Retkowski, Alexander Waibel

TL;DR

The paper addresses the lack of standardized paragraph segmentation benchmarks for speech transcripts by introducing TEDPara and YTSegPara and proposing a constrained decoding approach that allows LLMs to insert paragraph breaks without altering the transcript. It demonstrates that a compact model, MiniSeg, can jointly predict chapters and paragraphs in a hierarchical framework, with efficient section-wise processing and minimal performance trade-offs, especially when pretraining on related segmentation tasks. The authors provide comprehensive automatic and human evaluations, show that constrained decoding reduces hallucinations compared to unconstrained generation, and argue for treating paragraph segmentation as a practical, standardized task in speech processing with clear downstream benefits. Collectively, these contributions lay the groundwork for robust paragraph segmentation in speech transcription and broader text-segmentation research, enabling improved readability, retrieval, and downstream summarization.

Abstract

Automatic speech transcripts are often delivered as unstructured word streams that impede readability and repurposing. We recast paragraph segmentation as the missing structuring step and fill three gaps at the intersection of speech processing and text segmentation. First, we establish TEDPara (human-annotated TED talks) and YTSegPara (YouTube videos with synthetic labels) as the first benchmarks for the paragraph segmentation task. The benchmarks focus on the underexplored speech domain, where paragraph segmentation has traditionally not been part of post-processing, while also contributing to the wider text segmentation field, which still lacks robust and naturalistic benchmarks. Second, we propose a constrained-decoding formulation that lets large language models insert paragraph breaks while preserving the original transcript, enabling faithful, sentence-aligned evaluation. Third, we show that a compact model (MiniSeg) attains state-of-the-art accuracy and, when extended hierarchically, jointly predicts chapters and paragraphs with minimal computational cost. Together, our resources and methods establish paragraph segmentation as a standardized, practical task in speech processing.

Paragraph Segmentation Revisited: Towards a Standard Task for Structuring Speech

TL;DR

The paper addresses the lack of standardized paragraph segmentation benchmarks for speech transcripts by introducing TEDPara and YTSegPara and proposing a constrained decoding approach that allows LLMs to insert paragraph breaks without altering the transcript. It demonstrates that a compact model, MiniSeg, can jointly predict chapters and paragraphs in a hierarchical framework, with efficient section-wise processing and minimal performance trade-offs, especially when pretraining on related segmentation tasks. The authors provide comprehensive automatic and human evaluations, show that constrained decoding reduces hallucinations compared to unconstrained generation, and argue for treating paragraph segmentation as a practical, standardized task in speech processing with clear downstream benefits. Collectively, these contributions lay the groundwork for robust paragraph segmentation in speech transcription and broader text-segmentation research, enabling improved readability, retrieval, and downstream summarization.

Abstract

Automatic speech transcripts are often delivered as unstructured word streams that impede readability and repurposing. We recast paragraph segmentation as the missing structuring step and fill three gaps at the intersection of speech processing and text segmentation. First, we establish TEDPara (human-annotated TED talks) and YTSegPara (YouTube videos with synthetic labels) as the first benchmarks for the paragraph segmentation task. The benchmarks focus on the underexplored speech domain, where paragraph segmentation has traditionally not been part of post-processing, while also contributing to the wider text segmentation field, which still lacks robust and naturalistic benchmarks. Second, we propose a constrained-decoding formulation that lets large language models insert paragraph breaks while preserving the original transcript, enabling faithful, sentence-aligned evaluation. Third, we show that a compact model (MiniSeg) attains state-of-the-art accuracy and, when extended hierarchically, jointly predicts chapters and paragraphs with minimal computational cost. Together, our resources and methods establish paragraph segmentation as a standardized, practical task in speech processing.
Paper Structure (47 sections, 1 equation, 3 figures, 6 tables)

This paper contains 47 sections, 1 equation, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Paragraph Segmentation in Written Text.
  4. Segmentation of Speech Transcripts.
  5. Text Segmentation Benchmarks.
  6. Constrained Decoding with LLMs.
  7. Task Definition
  8. Dataset Construction
  9. TEDPara
  10. Preprocessing.
  11. Release.
  12. YTSegPara
  13. Augmentation.
  14. Utility and Scope.
  15. Release.
  16. ...and 32 more sections

Figures (3)

  • Figure 1: Paragraph segmentation turns an undifferentiated transcript into visually coherent paragraphs, aiding readability and navigation.
  • Figure 2: Conceptual illustration of the sentence-wise constrained decoding algorithm for paragraph insertion with LLMs. At each sentence boundary, the model performs a single forward pass to decide between two constrained actions: HTML]f49f55continue (emit standard punctuation) or HTML]bad8aabreak (emit punctuation followed by newlines).
  • Figure 3: Prompt Template for Paragraph Insertion with LLMs