A perceptual study on variability in break allocation within Chinese sentences

Min Chu, Honghui Dong, Jianhua Tao2006 · DOI

Abstract

Traditional Text-to-Speech (TTS) systems often operate under the assumption that there is a single, optimal "gold standard" for break allocation in any given sentence. This study challenges that rigid approach by investigating human preferences for various pause configurations. Using a two-step framework of chunking and breaking, we demonstrate that prosodic chunks serve as the atomic rhythmic units of Chinese sentences. Our findings indicate that while breaking inside these chunks significantly degrades naturalness, multiple break configurations are perceived as equally acceptable provided they respect chunk boundaries and maintain rhythmic length balance.

Introduction

In Text-to-Speech (TTS) synthesis, "break allocation"—the decision of where to insert pauses—is crucial for natural-sounding prosody. While algorithmic models often seek a single "correct" break location, human speech exhibits significant variability. This paper addresses the gap between rigid algorithmic predictions and flexible human perception. The authors investigate whether "prosodic chunks"—stable rhythmic units that speakers instinctively avoid splitting—serve as the fundamental building blocks of Chinese sentences, and whether breaks are governed more by rhythmic constraints than complex syntactic rules.

Methodology

The researchers conducted a perceptual experiment based on a two-step framework for rhythmic organization.

The two-step framework for rhythmic organization: Step 1 involves identifying prosodic chunks, and Step 2 involves grouping these chunks into phrases.

As illustrated in Figure 1, the process involves:

  1. Chunking: Identifying "prosodic chunks," which are words or groups of words that should not be split.
  2. Breaking: Grouping these chunks into larger prosodic phrases by inserting breaks.

The study utilized 20 Chinese sentences (13–19 syllables long) synthesized using the Mulan TTS system. For each sentence, eight different break versions were generated: six "positive" versions (breaks placed at chunk boundaries) and two "negative" versions (breaks placed intentionally inside a prosodic chunk). Twenty human subjects performed an AB test across 560 pairs of utterances to determine a Preference Rate (PR) for each strategy.

Results and Analysis

The experiment yielded two primary insights regarding the nature of Chinese prosody.

The Atomic Nature of Prosodic Chunks

The most critical finding is that prosodic chunks act as the "atoms" of Chinese rhythm. As shown in Figure 2, there is a stark contrast in listener preference based on break placement.

Comparison of Preference Rates (PR) between positive samples (breaks at chunk boundaries) and negative samples (breaks inside chunks).

Positive samples, where breaks occurred at chunk boundaries, were rated significantly higher than negative samples, where breaks split a chunk (p < 0.01). This confirms that respecting the integrity of these atomic units is the most vital factor for naturalness.

Flexibility and Rhythmic Balance

Once chunk boundaries were respected, listeners were surprisingly flexible. There was no statistically significant difference (p = 0.45) in preference between sentences with no breaks, one break, two breaks, or "natural" breaks copied from human recordings.

Furthermore, the study suggests that for sentences of this length, rhythmic constraints outweigh complex syntax. The authors introduced the Length Balance Ratio (LBR) and found a negative correlation (-0.44) between LBR and preference. This indicates that as the length difference between resulting phrases increases, perceived naturalness decreases. In single-break sentences, the most preferred versions were those where the break was placed closest to the center of the sentence.

Conclusion

This study demonstrates that there is no single "perfect" break for a Chinese sentence. Instead, TTS systems should focus on identifying "no-break" locations—the internal parts of prosodic chunks—and ensuring that any inserted breaks maintain a sense of rhythmic length balance. Future research may explore how these findings scale to longer, more syntactically complex sentences where structural rules might play a more dominant role.

Summary updated

BibTeX
@misc{chu2006perceptual,
  author = {Min Chu and Honghui Dong and Jianhua Tao},
  title = {{A perceptual study on variability in break allocation within Chinese sentences}},
  year = {2006},
  doi = {10.21437/speechprosody.2006-160},
}
Figure