Spoken Keyword Rescoring and Document Retrieval for Low-resource Languages

For languages that have adequate data for automatic speech recognition (ASR), many keyword search(KWS) and document retrieval(SDR) systems have been developed with near-optimal performance. However, lacking of sufficient training data to produce high accuracy transcript, identification and retrieval of queries in speech data from low-resources languages remains challenging. To compensate for these shortcomings, we extracted signals from the data that are useful in KWS/SDR. These signals take multiple forms word/morpheme burstiness, rescored confusion network posteriors, acoustic/prosodic qualities, phoneme recognition results, etc. Different strategies have been proposed: 1) a four-way classification of keyword hypotheses; 2) learning-to-rank algorithms; 3) fusion of symbolic subsequence matching and information-based dynamic time wrapping; 4) prosody-based language model. We evaluated these strategies to the speech data from IARPA-BABEL and MediaEval-QUESST Workshop. The paper summarizes the previous work of author in spoken keyword rescoring and document retrieval, as well as discusses the ongoing exploration.