Automatic learning of phonetic mappings for cross-language phonetic-search in keyword spotting

Abstract

Phonetic-search (PS) is an extremely fast technique used for spoken keyword spotting over large amounts of audio data. PS is based on matching a desired phonetic pattern over existing phonetic lattices, avoiding heavy computations of acoustic probabilities during the search. Since PS requires substantial acoustic and language resources (LR) for training acoustic models, there is a need for reducing model training costs to support new target languages. Particular cases of under-resourced languages pose even a greater challenge for PS as the available LR are not sufficient for acoustic model training. This study examines methods for keyword search in a new target language, using existing models of another source language in the lattice generation phase. We explore methodologies for learning cross-language phonetic mappings depending on the availability of data in the target language. We describe three approaches for creating phonetic-mappings: linguistic, acoustic, and statistic, introducing an efficient way for learning a robust statistical cross-language mapping. Our cross-language PS experiments showed that learning a good cross-language mapping can alleviate acoustic mismatches between languages, to significantly improve cross-language phonetic-search.