A Light-weight contextual spelling correction model for customizing transducer-based speech recognition systems