Abstract
Keyword spotting (KWS) systems often struggle with recognition errors and out-of-vocabulary (OOV) terms. This paper introduces a discriminative training approach for a phoneme confusion model designed to mitigate these issues. By directly optimizing the Figure of Merit (FOM) using a differentiable sigmoid approximation, the model learns to effectively expand search indices. Experimental results demonstrate that this approach significantly improves the trade-off between hit rates and false alarms, effectively halving the false alarm rate at specific operating points on the NIST STD06 evaluation set.
Introduction
Keyword spotting involves identifying specific terms within large audio datasets. Traditional systems rely on word-level indexing, which is vulnerable to two primary issues: acoustic recognition errors due to noise and the inability to index out-of-vocabulary (OOV) terms like names or technical jargon. While phoneme-level indexing offers a solution for OOVs, phoneme recognizers are inherently error-prone. This research proposes a "corrector" mechanism—a phoneme confusion model—that maps intended phonemes to potential misrecognitions in the index, thereby increasing the robustness of the search.
Methodology
The system represents both the audio index and search queries as Weighted Finite State Transducers (WFSTs). The core component is a Confusion Model, a specialized WFST that handles phoneme substitutions, insertions, and deletions.
Unlike standard approaches that build confusion models by counting errors in training data, this paper employs Discriminative Training. The weights of the confusion model are treated as parameters optimized against the Figure of Merit (FOM). Because the FOM is a non-differentiable step function, the authors utilize a sigmoid approximation to enable gradient descent. This allows the model to prioritize correcting confusions that most significantly impact search performance.
Experimental Setup
The acoustic models were trained on a 3,000-hour corpus of English broadcast news. The confusion model itself was trained using 8 hours of Broadcast News data with 1,217 training queries. Evaluation was conducted on the NIST STD06 set, comprising 3 hours of audio and 1,104 test queries.
Results and Discussion
The discriminative training process showed substantial gains during the optimization phase. As the number of iterations increased, the model's ability to rank true hits above false alarms improved significantly, as seen in the training data performance curves.
A key finding was the necessity of "smoothing" the trained weights. A raw trained model can overfit, leading to performance degradation at low false alarm rates. By interpolating the trained model with a "null model" (which assumes perfect recognition), the authors achieved a more robust balance. On the NIST evaluation set, this smoothed model consistently outperformed the baseline.
Quantitatively, for a target hit rate of 0.45, the trained model reduced the false alarm rate from 5 to 2.5 per hour per query—a 50% improvement.
Conclusion and Future Work
The study demonstrates that direct optimization of search metrics like FOM is a powerful technique for improving KWS performance. Future work will explore context-dependent (bigram) confusion models to capture more nuanced phonetic errors. Additionally, integrating this phoneme-level model into hybrid word-phoneme systems and operating directly on acoustic scores could further enhance system robustness.
Summary updated