Abstract
Modern keyword search (KWS) systems often rely on Large Vocabulary Continuous Speech Recognition (LVCSR) engines trained to minimize overall Word Error Rate (WER). However, KWS tasks prioritize the accurate detection of specific target terms over general transcription accuracy. This paper introduces a "Keyword-Boosted" state-level Minimum Bayes Risk (sMBR) training criterion for Deep Neural Network (DNN) acoustic models. By weighting keyword-related states more heavily during training, the proposed method achieves a relative performance gain of up to 6.1% in keyword search accuracy on low-resource Vietnamese speech data.
Introduction
The standard training objective for acoustic models in LVCSR systems is the state-level Minimum Bayes Risk (sMBR) criterion, which aims to minimize expected state-level errors. While effective for reducing WER, sMBR treats all lexical units with equal importance. In KWS applications, the primary goal is to identify specific "needles" in the audio haystack. Misidentifying common filler words is often inconsequential compared to missing a critical keyword. This research addresses the mismatch between general-purpose training objectives and the specific requirements of keyword spotting.
Methodology: Keyword-Boosted sMBR
The authors propose a Keyword-Boosted sMBR (KW-boosted sMBR) criterion to redistribute the model's focus during training.
- Standard sMBR: The DNN is trained by minimizing expected state-level errors across utterances using a State Accuracy function (assigning 1 for correct states, 0 otherwise).
- Boosting Mechanism: A boosting weight, $\alpha$, is introduced into the State Accuracy function. States belonging to a keyword in the reference transcript have their accuracy scores multiplied by $\alpha$, while non-keyword states retain a weight of 1.
- Implementation: This modification is computationally lightweight as it only adjusts weights within the existing objective function without increasing algorithmic complexity. This encourages the DNN to capture finer acoustic details for target keyword states.
Experimental Setup
The method was evaluated using the IARPA Babel program’s Vietnamese limited-language-pack (LLP), consisting of 10 hours of transcribed conversational telephone speech for training and a 15-hour evaluation set (evalpart1). Two keyword lists were utilized: a large evaluation list of 4,065 phrases and a smaller development list of 200 phrases.
Results and Discussion
The experimental results highlight a fundamental trade-off between general transcription accuracy and keyword detection performance.
Precision-Recall and Metric Divergence
Boosting keywords increases the system's sensitivity, leading to higher recall (fewer deletions) but also a higher rate of false alarms (insertions). Interestingly, while the Actual Term Weighted Value (ATWV)—the primary KWS metric—improved significantly, the overall Word Error Rate (WER) slightly increased. This confirms that optimizing for specific keywords is a distinct task from optimizing for general transcription.
Sensitivity to Hyperparameters
The optimal boosting weight $\alpha$ depends on the keyword list size. As shown in Figure 2, for the large evaluation list, a moderate boost ($\alpha \approx 5$) was optimal. For smaller lists, performance was less stable, suggesting potential overfitting to specific acoustic patterns.
Comparative Performance
The KW-boosted sMBR system consistently outperformed the conventional sMBR baseline. Across various test configurations, including those with improved lexicons (G2P) and specialized Language Models (CS-KWLM), the gains from acoustic boosting were additive.
As shown in the table above, the most advanced configuration reached an ATWV of 0.3396, representing a significant improvement over the standard sMBR baseline. Analysis of the 1-best output revealed that the boosted system reduced keyword deletions by approximately 33%.
Conclusion
The Keyword-Boosted sMBR criterion provides a targeted approach to enhancing acoustic models for KWS. While it successfully reduces missed detections, it also increases false alarms. Future work will explore "symmetric" boosting weights to explicitly penalize false alarms in decoding lattices, potentially balancing the precision-recall trade-off more effectively.
Summary updated