Abstract
Prosody labeling is a critical yet labor-intensive component of high-quality Text-to-Speech (TTS) synthesis. This paper introduces an automated framework for generating Korean Tone and Break Index (K-ToBI) labels using lexico-syntactic features and decision tree classifiers. By automating the labeling of large-scale corpora, the system enables the production of natural-sounding Korean speech without the prohibitive costs of manual linguistic annotation.
Introduction
In speech synthesis, prosody—the rhythm, stress, and intonation of speech—is essential for naturalness. The Tone and Break Index (ToBI) framework provides a symbolic representation of these features. However, creating a ToBI-labeled corpus for Korean is challenging due to the language's unique prosodic structure, which emphasizes phrasing over lexical stress. This study addresses the bottleneck of manual labeling by proposing a system for automatic K-ToBI label generation and prediction.
The K-ToBI Representation
Korean prosody is hierarchically organized into Intonational Phrases (IP) and Accentual Phrases (AP). Unlike English, which uses pitch accents on specific syllables, Korean intonation is primarily defined by boundary tones at the ends of these phrases. This structure allows for a symbolic representation that can be mapped directly to acoustic features.
Methodology
The proposed system utilizes a multi-stage architecture to transition from raw speech to symbolic prosodic labels and finally to synthesized pitch contours.
1. Automatic K-ToBI Labeling
The system processes raw audio by extracting pitch sequences and aligning them with manually constructed phone scripts. It segments the speech into phrase units and analyzes pitch characteristics (min, max, and average) to assign potential pitch patterns (High, Low, or Middle). These are then mapped to standard K-ToBI symbols.
2. Tone Label Prediction
Using the automatically labeled corpus, a C4.5-style decision tree is trained to predict tones from text. The model incorporates lexico-syntactic features such as Part-of-Speech (POS) tags, syllable structure (onset/nucleus/coda), and the word's position within a phrase. Syllable structure is particularly important in Korean, as glottal sounds often correlate with higher tones.
3. F0 Contour Generation
Finally, a linear regression model converts the predicted symbolic labels into a fundamental frequency (F0) contour, which serves as the melodic blueprint for the synthesized voice.
Dataset and Experimental Setup
The research utilized the MBCNEWSDB, a corpus of 10,149 Korean news broadcast sentences (approximately 122,047 words). The dataset was split into 8,780 training and 1,369 testing sentences. Given the formal nature of news broadcasts, the intonation is relatively constrained, with a high frequency of "L%" (low boundary) tones.
Results
The automated system demonstrated competitive performance compared to manually labeled benchmarks:
- Boundary Tone Prediction: The decision tree achieved an error rate of 16.10% for Intonation Phrase (IP) boundary tones after pruning.
- Accentual Phrase Prediction: Internal phrase melody prediction resulted in a 30.40% error rate.
- F0 Generation Accuracy: The final pitch contours achieved a Root Mean Squared Error (RMSE) of 22.219 Hz and a correlation coefficient of 0.595.
Discussion and Conclusion
The study highlights the effectiveness of using K-ToBI as a modular intermediate layer in TTS systems. While the current model performs well on news-style data, its reliance on a specific domain suggests potential limitations in handling emotional or casual dialogue. Future work will focus on expanding the methodology to more diverse datasets and addressing data skew issues inherent in specific speaking styles. By simplifying the labeling process without sacrificing naturalness, this approach paves the way for more efficient development of high-quality, language-specific TTS systems.
Summary updated