Phonotactic spoken language recognition: Using diversely adapted acoustic models in parallel phone recognizers

摘要

In phonotactic spoken language recognition systems, acoustic model adaptation prior to phone lattice decoding has been adopted to deal with the mismatch between training and test conditions. Moreover, combining diversified phonotactic features is commonly used. These motivate us to have an in-depth investigation of combining diversified phonotactic features from diversely adapted acoustic models. Our experiment shows that our approach achieves an equal error rate (EER) of 1.94% in the 30-second closed-set trials of the 2007 NIST Language Recognition Evaluation (LRE). It represents a 14.9% relative improvement in EER over a sophisticated system, in which parallel phone recognizers, speaker adaptive training (SAT) in acoustic models and CMLLR adaptation are used. Moreover, it is shown that our approach provides consistent and substantial improvements in three different phonotactic systems, in each of which a single phone recognizer is used.