This paper identifies and models a phenomenon observed across low-resource languages in keyword search results from speech retrieval systems where the speech recognition has high error rate, due to very limited training data. High confidence correct detections (HCCDs) of keywords are rare, yet often succeed one another closely in time. We refer to these close sequences of HCCDs as keyword hotspots. The ability to predict keyword hotspots could support speech retrieval, and provide new insights into the behavior of speech recognition systems. We treat hotspot prediction as a binary classification task on all word-sized time intervals in an audio file of a telephone conversation, using prosodic features as predictors. Rare events that follow this pattern are often modeled as a self-exciting point process (SEPP), meaning the occurrence of a rare event excites a following one. To label successive points in time as occurring within a hotspot or not, we fit a SEPP function to the distribution of HCCDs in the keyword search output. Two major learning challenges are that the size of the positive class is very small, and the training and test data have dissimilar distributions. To address these challenges, we develop a novel data selection framework that chooses training data with good generalization properties. Results exhibit superior generalization performance.
All Science Journal Classification (ASJC) codes
- Theoretical Computer Science
- Human-Computer Interaction