Introduction
In HMM-based text-to-speech (TTS) synthesis, the quality of generated speech depends heavily on linguistic context. Unlike speech recognition, which typically uses triphones, synthesis systems track dozens of features—such as syllable stress, phrase position, and part-of-speech—to capture prosodic nuances. However, this leads to a "combinatorial explosion." With 26 features, there are theoretically $10^{27}$ possible unique states, creating a massive, sparse model space that is difficult to train effectively with limited data.
Bayesian Networks for Feature Selection
To address this redundancy, the authors propose using Bayesian Networks (BNs) to discover the underlying statistical dependencies between linguistic context and acoustic output. By learning the structure of these networks, it becomes possible to identify which features directly influence acoustics and which are conditionally independent.
The methodology involves three primary steps:
- Quantization: Continuous acoustic data (Mel-cepstrum, F0, and duration) is converted into discrete classes using Vector Quantization (VQ) to allow for standard BN structure learning.
- Structure Learning: Algorithms like LAGD Hill Climbing and Tree Augmented Naive Bayes (TAN) are employed to map dependencies between context features and acoustic nodes.
- Feature Pruning: Features that are distant or disconnected from the acoustic nodes in the learned BN are deemed redundant and discarded.
Experimental Results
The study utilized a British English corpus of 2,969 manually labeled utterances from a single male speaker. The BN analysis revealed significant redundancy in standard feature sets. For instance, while many features describe phrase structure, only a subset directly impacts pitch (F0) or duration. For spectral features, phoneme identity and its immediate neighbors remained the most critical factors.
The authors successfully reduced the context feature set from 26 down to just 6 for spectral modeling and 9 for both F0 and duration modeling. To validate this reduction, they compared a "topline" system using all 26 features against a "reduced" system in a forced-choice subjective listening test with 27 native speakers.
The results showed that the full system was preferred only 55% of the time, compared to 45% for the reduced system. This negligible difference indicates that approximately 70% of the standard context features are unnecessary for maintaining high-quality speech synthesis.
Conclusion
This research demonstrates that the "data sparsity problem" in TTS is largely self-inflicted by the inclusion of irrelevant features. By pruning the feature set using Bayesian Networks, researchers can create denser, more manageable models without sacrificing perceptible quality. Future work aims to automate this selection process and potentially replace traditional decision-tree clustering with direct BN-based acoustic modeling to better handle the sparsely populated context space.
Summary updated