Abstract
This paper addresses the performance gap between in-vocabulary and out-of-vocabulary (OOV) queries in Chinese Spoken Term Detection (STD). We propose a two-stage pipeline: partial phonetic matching to maximize recall, followed by a novel simulated lattice rescoring method to enhance precision. Experimental results on a 104-hour spontaneous speech corpus demonstrate a 24.1% relative improvement in OOV search accuracy, measured by Figure of Merit (FOM).
Introduction
Spoken Term Detection (STD) is the task of finding specific keywords within large archives of audio. Most modern systems rely on Large Vocabulary Continuous Speech Recognition (LVCSR) to convert audio into word lattices—graphs representing possible word sequences. The primary challenge for these systems is the Out-Of-Vocabulary (OOV) problem, where search terms missing from the recognizer's dictionary lead to significant performance degradation.
In Chinese, it is often assumed that OOV issues are mitigated because words are composed of characters that are almost always present in the dictionary. However, this study debunks that myth. Matching an OOV word as a sequence of individual characters is highly inefficient, leading to a massive performance drop (e.g., 75.8% accuracy for in-vocabulary words vs. 59.1% for OOV words) due to the loss of probabilistic weight in the lattice structure.
Methodology
The authors propose a two-stage pipeline designed to balance the trade-off between finding every possible match (recall) and ensuring those matches are correct (precision).
Stage 1: Partial Phonetic Matching
To ensure high recall, the system converts the standard word lattice into a phonetic lattice based on "toneless half-syllables." Instead of exact word matches, it employs a loose matching criterion (a threshold of 0.7). This allows the system to flag potential candidates even if the phonetic sequence isn't a perfect match, preventing OOV terms from being missed due to minor recognition errors.
Stage 2: Lattice Rescoring
Phonetic matching often generates numerous false alarms. To filter these, the system "simulates" what would have happened if the OOV word had been in the dictionary originally. The OOV query is inserted as a new arc into the original word lattice at the candidate time intervals.
A new posterior probability for this arc is calculated by combining the acoustic score (how well the audio fits the OOV word's sounds) and a simplified language model score. This allows the system to rank candidates with precision comparable to full re-recognition but at a fraction of the computational cost.
Experimental Results
The system was evaluated on a 104-hour corpus of spontaneous Chinese speech, with acoustic models trained on 500 hours of speech. The query set included 512 OOV words and 1,104 In-Vocabulary (INV) words filtered from real-world logs.
Performance was measured using the Figure of Merit (FOM), a standard measure for keyword spotting.
- Baseline OOV Performance: 43.3% FOM.
- Phonetic Match Only: Increased recall but dropped FOM to 34.2% due to excessive false alarms.
- Full Two-Stage Method: Achieved 53.7% FOM, representing a 24.1% relative improvement over the baseline for OOV queries.
Conclusion
The proposed two-stage architecture effectively bridges the gap between OOV and INV performance in Chinese STD. By combining high-recall phonetic matching with high-precision lattice rescoring, the system achieves results close to an "Oracle" scenario. Future work will focus on improving the computational efficiency of acoustic scoring and integrating more sophisticated language model context to further refine precision.
Summary updated