Introduction
Standard Text-to-Speech (TTS) systems often struggle with naturalness due to a "one-size-fits-all" approach to phrasing. Phrasing—the placement of pauses in speech—is a critical component of prosody that varies significantly across different speaking styles. While traditional models rely on local context like part-of-speech (POS) n-grams, they often fail to capture the hierarchical structures and stylistic nuances inherent in human speech. This paper presents a novel method for learning style-specific phrasing by training Stochastic Context-Free Grammars (SCFGs) on prosodic structures, leading to more natural synthesized speech across diverse genres.
Methodology
The proposed approach to Phrase Break Prediction (PBP) involves two primary stages:
- Prosodic Parsing with SCFGs: Rather than using standard linguistic syntax, the authors define "prosodic constituents" based on actual pauses (breaks longer than 80ms). An SCFG is trained to learn the patterns of these constituents using POS tags as terminal symbols. This allows the model to capture the hierarchical "prosodic chunking" unique to a specific speaker or style.
- Feature Integration via CART Trees: Structural information derived from the SCFG—such as the count of open or closed brackets at word boundaries—is extracted as features. These hierarchical features are combined with standard local features (e.g., punctuation, POS tags) and fed into a Classification and Regression Tree (CART) to make the final binary decision: Break (B) or Non-Break (NB).
Style-Specific Data and Over-Pausing
The study evaluated five distinct corpora: ARCTIC-A (phonetically balanced), Europarl (formal parliamentary), F2B (radio news), Obama (rhetorical public address), and Emma (narrative audiobook). A key observation was that generic models tend to significantly over-predict pauses when applied to specialized styles.
As shown in the data above, the default prediction rate consistently exceeds the actual break frequency across all datasets. For instance, in the ARCTIC-A corpus, the default model predicted breaks at 8.96% of boundaries, whereas the actual speaker paused only 6.25% of the time. This discrepancy highlights the necessity of style-aware modeling to avoid robotic, over-segmented speech.
Results and Discussion
The style-specific hybrid models outperformed the baseline in both objective and subjective metrics. Objective accuracy (F1 score) improved across the board, with the most notable gain in the Emma audiobook corpus, rising from 69.98 to 82.94. Furthermore, Mel-Cepstral Distortion (MCD) decreased, indicating that the synthesized audio more closely matched the original recordings.
Subjective listening tests further validated these findings. On the F2B radio news corpus, listeners were asked to choose between the baseline and the new grammar-based model.
The results indicate a strong preference for the new model, with 72 participants favoring the grammar-based approach compared to 50 for the baseline. This suggests that capturing hierarchical prosodic structure significantly enhances the perceived naturalness and rhythmic "fingerprint" of the speech.
Conclusion
By shifting from a generic, local-context model to a style-specific, hierarchical grammar-based approach, the authors demonstrate a significant improvement in TTS phrasing. While the current work focuses on pause location, future research could extend this to pause duration and applications in machine translation to further improve the intelligibility and stylistic fidelity of synthesized speech.
Summary updated