Introduction
Traditional Large Vocabulary Continuous Speech Recognition (LVCSR) systems face a significant "brick wall": they are inherently dictionary-bound. If a user searches for a keyword that was not included in the system’s training vocabulary—an Out-of-Vocabulary (OOV) word—the system is effectively blind to it. This paper, presented at INTERSPEECH 2012 by researchers from Raytheon BBN Technologies, proposes a novel subword-based approach to overcome this limitation.
The Subword Trade-off
To detect unseen words, researchers typically replace full-word models with subword units like phones or syllables. However, this introduces a fundamental trade-off. Longer units, such as multigrams, provide richer acoustic context and better recognition of common patterns. Conversely, shorter units, like single phones, are more general and better suited for piecing together brand-new, unseen phonetic sequences. While existing methods attempt to find an "optimal" segmentation of training data into these units, this paper argues that such a singular choice is limiting.
Methodology: Overlapping Subword Lattices
The core innovation of this research is a fixed-length phonetic unit recognizer that utilizes flexible segmentation. Instead of relying on complex algorithms to determine the best subword boundaries, the authors utilize all possible substrings of one, two, and three phones found in the training data.
Rather than forcing training data into a single sequence of units, the system represents training utterances as lattices. This allows the model to accommodate overlapping units simultaneously—for example, a three-phone unit can be modeled alongside the three individual one-phone units that comprise it. This approach maximizes the utility of training data and reduces data fragmentation. For keyword search, recognition lattices are converted into consensus networks, and a Dynamic Programming (DP) algorithm is used for approximate phonetic matching, allowing for minor misrecognitions during the search process.
Experimental Setup
The system was trained on 100 hours of the Fisher Spanish conversational telephone speech (CTS) corpus and evaluated using 4 hours of the Spanish CallHome (Eval97) test set. The researchers categorized search terms into groups to test varying difficulty levels, including rare in-vocabulary words (IV-rare), words in the dictionary but absent from training (IV-OOT), and true OOV words.
Results and Analysis
The proposed fixed-length unit system was compared against standard word-based LVCSR, context-dependent phones, graphones, and variable-length phonetic units.
On truly unseen OOV words, the proposed fixed-length units achieved the highest Optimal Term Weighted Value (OTWV) of 0.595, significantly outperforming graphones (0.570) and variable-length units (0.545). Interestingly, while the word-based system maintained a lower Phone Error Rate (PER) of 29.97% compared to the fixed-length system's 33.55%, the latter was far more effective at finding OOV keywords. This proves that lower PER does not always correlate with superior search performance for unseen terms.
Conclusion
The results suggest that a rigid but overlapping subword structure is superior at generalizing to completely new phonetic sequences compared to algorithms that attempt to learn "optimal" variable-length chunks. Future work may involve exploring longer unit lengths and combining different subword models into ensemble systems to further improve detection accuracy.
Summary updated