Low-resource spoken keyword search strategies in georgian inspired by distinctive feature theory

We present low-resource spoken keyword search (KWS) strategies guided by distinctive feature theory in linguistics to conduct data selection, feature selection, and transcription augmentation. These strategies were employed in the context of the 2016 NIST Open Keyword Search Evaluation (OpenKWS16) using conversational Georgian from the IARPA Babel program. In particular, we elaborate on the following: (1) We exploit glottal-source-related acoustic features that characterize Georgian ejective phonemes ([+constricted glottis], [+raised larynx ejective] specified in distinctive feature theory). These features complement standard acoustic features, leading to a relative fusion gain of 11.9%. (2) We use noisy channel models to incorporate probabilistic phonetic transcriptions from mismatched crowdsourcing to conduct transfer learning to improve KWS for extremely under-resourced conditions (24 min of transcribed Georgian), achieving a relative improvement of 118% over the baseline and a relative fusion gain of 32%.(3) Using distinctive feature analysis, we select a compact subset of source languages used in past evaluations to ensure high phonetic coverage for cross-lingual acoustic modeling when only limited system development time and computational resources are available. This strategy leads to comparable performance to using all available linguistic resources when only 1/3 of the source languages were chosen.