We present a novel approach to building a subword speech recognizer for the task of phonetic keyword search. The recognizer, which uses short fixed-length phonetic units, is trained with phonetic transcripts that are segmented into all possible substrings of 1, 2 and 3 phones, using a lattice representation to accommodate the overlapping units. We compare the keyword search accuracy of the proposed system with systems that use words, graphones, variable-length phonetic units and context-dependent phones. Experiments with Spanish CTS data show that the proposed subword recognizer outperforms other subword systems in terms of phonetic keyword search accuracy measured on queries that consist of words not present in the training data.

Index Terms: speech recognition, keyword search, OOV