Combined Use of Speaker- and Tone-Normalized Pitch Reset with Pause Duration for Automatic Story Segmentation in Mandarin Broadcast News

Lei Xie, Chuan Liu, Helen MengNAACL 2007 · OpenReview

Abstract

This paper investigates the combined use of pause duration and pitch reset for automatic story segmentation in Mandarin broadcast news. Analysis shows that story boundaries cannot be clearly discriminated from utterance boundaries by speaker-normalized pitch reset due to its large variations across different syllable tone pairs. Instead, speaker- and tone-normalized pitch reset can provide a clear separation between utterance and story boundaries. Experiments using decision trees for story boundary detection reinforce that raw and speaker-normalized pitch resets are not effective for Mandarin Chinese story segmentation. Speaker- and tone-normalized pitch reset is a good story boundary indicator. When it is combined with pause duration, a high F-measure of 86.7% is achieved. Analysis of the decision tree uncovered four major heuristics that show how speakers jointly utilize pause duration and pitch reset to separate speech into stories.

The Challenge: Tonality vs. Structure

In automatic broadcast news processing, "story segmentation" is the task of dividing a continuous audio stream into individual news items. In languages like English, speakers typically signal a new story using two main prosodic cues: long pauses and a "pitch reset" (starting the new sentence at a significantly higher pitch than the end of the previous one).

However, Mandarin is a tonal language. The pitch of a syllable isn't just for structural emphasis; it defines the word itself. Because different tone combinations (e.g., a "low" tone followed by a "high" tone) naturally create pitch jumps, the structural "pitch reset" used to mark a new story gets buried in the lexical "noise" of the tones. Standard speaker normalization helps, but it isn't enough to reveal the underlying story structure in Mandarin.

Figure 3: (a) Illustration of pitch reset between speech units. (b) The four basic tones in Mandarin Chinese, which complicate structural pitch analysis.

Methodology: Multi-Stage Normalization

The authors propose a multi-stage normalization process to isolate the structural pitch reset from tonal interference:

  1. Speaker Normalization: A standard z-score normalization accounts for different natural pitch ranges (e.g., male vs. female).
  2. Tone Normalization: This core innovation involves calculating the mean and standard deviation for each of the four Mandarin tones across the corpus. Boundary syllables are then normalized based on their specific tone, effectively asking if a pitch is high relative to the expected range for that specific tone.
  3. Region of Interest (ROI): The analysis focuses specifically on the last syllable of one utterance and the first syllable of the next, separated by a pause.
  4. Decision Tree Classification: A C4.5 decision tree combines these normalized pitch features with pause duration to validate candidate story boundaries.

Key Results and Insights

The performance gain from tone normalization is substantial. Without it, pitch reset values vary wildly across different tone pairs, making it impossible to set a single threshold for a story boundary. Once tone normalization is applied, a clear statistical separation emerges between simple utterance boundaries and actual story boundaries across all tone combinations.

Figure 5: Mean speaker- and tone-normalized pitch reset across different tone pairs, showing a clear separation (shaded region) between utterance and story boundaries.

As shown in the table below, while pause duration is the strongest single indicator, adding the tone-normalized pitch reset significantly reduces errors:

Feature Set F-Measure
Raw Pitch Reset 50.8%
Speaker-Normalized Pitch Reset 55.3%
Speaker- & Tone-Normalized Pitch Reset 71.1%
Pause Duration alone 82.2%
Pause + Tone-Normalized Pitch Reset 86.7%

Analysis of the decision tree revealed how speakers balance these cues. For example, some reporters use shorter pauses but compensate with a larger pitch reset to signal a new story. Conversely, the model learned that a long pause without a significant pitch reset is likely a false alarm, such as a broadcast interruption.

Conclusion

This research demonstrates that while tonality masks structural cues in Mandarin, appropriate normalization can recover them for effective story segmentation. Future work will explore more complex tonal systems in other dialects and integrate lexical cues (e.g., "Coming up next...") to further improve accuracy.

Summary updated

Abstract

This paper investigates the combined use of pause duration and pitch reset for automatic story segmentation in Mandarin broadcast news. Analysis shows that story boundaries cannot be clearly discriminated from utterance boundaries by speaker-normalized pitch reset due to its large variations across different syllable tone pairs. Instead, speaker- and tone-normalized pitch reset can provide a clear separation between utterance and story boundaries. Experiments using decision trees for story boundary detection reinforce that raw and speaker-normalized pitch resets are not effective for Mandarin Chinese story segmentation. Speaker- and tone-normalized pitch reset is a good story boundary indicator. When it is combined with pause duration, a high F-measure of 86.7% is achieved. Analysis of the decision tree uncovered four major heuristics that show how speakers jointly utilize pause duration and pitch reset to separate speech into stories.
BibTeX
@inproceedings{xie2007combined,
  author = {Lei Xie and Chuan Liu and Helen Meng},
  title = {{Combined Use of Speaker- and Tone-Normalized Pitch Reset with Pause Duration for Automatic Story Segmentation in Mandarin Broadcast News}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2007},
  url = {https://openreview.net/forum?id=ryZfaXWOZS},
}
Figure