Introduction
In Mandarin Text-to-Speech (TTS) systems, the naturalness of synthesized speech is heavily dependent on accurate prosodic phrase prediction. This process determines where pauses should occur within a sentence to reflect human-like rhythm and intonation. Despite advancements, approximately 40% of text analysis errors in Mandarin TTS are attributed to incorrect prosodic boundary placement. These errors often stem from two primary challenges: the "aging" of training data relative to modern language use and the violation of grammatical (syntactic) constraints, which results in "extremely unacceptable" speech for listeners.
Methods
To address these issues, this research proposes a dual-optimization strategy involving unsupervised domain adaptation and the integration of structural syntactic features.
Unsupervised Adaptation via Punctuation
Manually labeling large-scale corpora for prosodic boundaries is resource-intensive. However, punctuation marks like commas often serve as reliable indicators of prosodic breaks. The authors utilized a massive pool of unlabelled modern web text (over 600 million words) to adapt their models. To ensure the quality of the "free" labels provided by commas, a dual-model filtering mechanism was employed. A sentence was only added to the training set if both a Maximum Entropy (ME) model and a Conditional Random Field (CRF) model agreed that a comma position represented a valid prosodic boundary. This approach allows the system to learn contemporary vocabulary and phrasing patterns without manual intervention.
Syntactic Feature Extraction
To prevent pauses that disrupt the grammatical logic of a sentence, the researchers integrated features derived from a syntactic parser (the Stanford Parser). By analyzing the hierarchical structure of a sentence, the model can determine the "tightness" between adjacent words.
The extracted features include Path Length (distance between words in the syntax tree), Sub-path Length (distance to the lowest common ancestor), and Sub-paths (the sequence of grammatical tags connecting words).
Experimental Results
The proposed methods were evaluated using a baseline trained on the People’s Daily corpus (1993–1996) and tested on both matched data and modern web news (2009).
The combination of unsupervised adaptation and syntactic features yielded a significant reduction in the Sentence Unacceptable Rate (SUR), which dropped from 15.9% to 8.6%—a relative improvement of 45%. Subjective evaluations also showed an increase in naturalness scores from 4.14 to 4.30 on a 5-point scale. Objective metrics on the matched dataset confirmed that syntactic features alone improved the F-score from 61.83% to 63.55%.
Conclusion
This study demonstrates that leveraging unlabelled web data through comma-based adaptation and incorporating structural syntactic constraints can substantially improve the naturalness of Mandarin TTS. While the system's performance is currently tied to the accuracy of the underlying syntactic parser, the results highlight the importance of bridging the gap between statistical word sequences and deep grammatical structures for high-quality speech synthesis.
Summary updated