Hybrid word-subword decoding for spoken term detection

This paper deals with a hybrid word-subword recognition system for spoken term detection. The decoding is driven by a hybrid recognition network and the decoder directly produces hybrid word-subword lattices. One phone and two multigram models were tested to represent sub-word units. The systems were evaluated in terms of spoken term detection accuracy and the size of index. We concluded that the best subword model for hybrid word-subword recognition is the multigram model trained on the word recognizer vocabulary. We achieved an improvement in word recognition accuracy, and in spoken term detection accuracy when in-vocabulary and out-of-vocabulary terms are searched separately. Spoken term detection accuracy with the full (in-vocabulary and out-of-vocabulary) term set was slightly worse but the required index size was significantly reduced.