Summarizing speech without text using hidden Markov models

Sameer Maskey, Julia Hirschberg

2006 · 75 citations · 11 references

DOIFull text

Open access

Concepts

TL;DR

The study proposes a text‑independent summarization method for speech documents using a Hidden Markov Model framework. The approach models sentence inclusion as hidden states, uses acoustic/prosodic features as observations, and selects the optimal sequence of segments to form a summary, which is then compared to human‑generated summaries. Results show that acoustic/prosodic cues alone can produce satisfactory summaries, demonstrating the feasibility of text‑independent summarization for spoken documents.

Abstract

We present a method for summarizing speech documents without using any type of transcript/text in a Hidden Markov Model framework. The hidden variables or states in the model represent whether a sentence is to be included in a summary or not, and the acoustic/prosodic features are the observation vectors. The model predicts the optimal sequence of segments that best summarize the document. We evaluate our method by comparing the predicted summary with one generated by a human summarizer. Our results indicate that we can generate 'good' summaries even when using only acoustic/prosodic information, which points toward the possibility of text-independent summarization for spoken documents.

References

11