Introduction
In large-vocabulary continuous speech recognition (LVCSR), Weighted Finite State Transducers (WFSTs) are the standard framework for decoding. The decoding graph is typically formed by the composition $C \circ L \circ G$, where $G$ represents the language model, $L$ the lexicon, and $C$ the context-dependency transducer. As phonetic contexts grow more complex—moving from triphones to quinphones or incorporating features like word boundaries—the $C$ transducer often becomes a significant bottleneck due to its sheer size.
Traditionally, this is managed by building a phonetic decision tree to tie parameters and then compiling that tree into an FST. However, this two-step process is often inefficient because the tree-building phase is "blind" to the final transducer size. A split that yields a marginal gain in acoustic likelihood can lead to an exponential increase in the number of states in the resulting transducer.
Methodology: Direct Construction with Regularization
The authors propose a novel approach: constructing the context-dependency transducer directly from training data using a size-aware regularization term. By integrating transducer construction into the greedy optimization process used for splitting phonetic contexts, the algorithm can explicitly manage the trade-off between model accuracy and decoder efficiency.
The Objective Function
The core of this method is a modified objective function used during the context-splitting phase: $$L(t) = G(t) - \alpha \cdot S(t)$$ Where:
- $G(t)$ is the gain in acoustic likelihood from the split.
- $S(t)$ is the number of new states required in the $C$ transducer to accommodate the split.
- $\alpha$ is a regularization weight.
When $\alpha = 0$, the method behaves like a standard decision tree construction. Increasing $\alpha$ penalizes the creation of new states, favoring more compact transducers.
Iterative State Splitting
The algorithm begins with a basic monophone model and iteratively applies splits. Throughout this process, the transducer is maintained as a minimal deterministic automaton. This ensures that $S(t)$ accurately reflects the incremental complexity added by each split.
Experimental Evaluation
The method was tested on a large-scale Google in-house spoken-query task, utilizing 2100 hours of training data and a 1-million-word vocabulary. The evaluation focused on how different values of $\alpha$ affected both the transducer size (number of states) and the Word Error Rate (WER).
Key Results
The experiments demonstrated that substantial reductions in model size are possible with minimal impact on accuracy. For 4-gram models ($n=4$), the conventional construction resulted in 79,507 states. By applying the proposed method with $\alpha=1000$, the number of states was reduced to just 6,302—a 12.6x reduction—while maintaining a WER of 21.6%.
For triphone models ($n=3$), the method reduced the state count from 1,849 to 1,032 ($\alpha=100$) and even showed a slight improvement in WER (from 21.2% to 21.1%). Furthermore, when incorporating word boundary features—a process that typically causes significant state explosion—the method achieved a 75% reduction in transducer size compared to the baseline.
Discussion and Insights
The primary insight of this work is that transducer size should be treated as a "first-class citizen" during model training. The authors observed that many splits in traditional decision trees provide negligible acoustic gains but carry high "costs" in terms of decoder complexity. By identifying and avoiding these expensive splits, the decoder remains efficient without sacrificing performance.
The framework is also highly flexible, easily extending to generalized features like vowel stress or word boundaries. Because it creates new states only when they are strictly necessary to distinguish models, it avoids the redundant state creation inherent in traditional methods that require separate phone sets for every feature variant.
Conclusion
Direct construction of context-dependency transducers with size-aware regularization offers a powerful solution to the complexity bottlenecks in LVCSR. By bridging the gap between acoustic modeling and FST compilation, this approach enables the use of richer phonetic contexts and additional features in real-time decoding environments. Future work may explore optimizing for the number of transitions (arcs) rather than just states to further refine decoder efficiency.
Summary updated