Small-vocabulary speech recognition for resource-scarce languages

We describe a technique for attaining high-accuracy, small-vocabulary speech recognition capability in resource-scarce languages that requires minimal audio data collection and no speech technology expertise. We start with an off-the-shelf commercial speech recognizer that has been trained extensively on a resource-rich language such as English. We then derive phonemic representations for any desired word in any target language, by a process of cross-language phonemic mapping. We show that this results in high accuracy recognition of vocabularies of up to several dozen words -- enough for many development-related applications such as information access, data collection, and simple transactions.