A low complexity cluster model interpolation based on-line adaptation technique for spoken query systems

The work presented in this paper describes the issues of on-line adaption in context of spoken query systems. In such systems, the available adaptation data is extremely small (≤ 3 seconds). Consequently, adapting such systems becomes extremely challenging. Moreover, since these systems are meant for real-time applications, the employed adaptation technique should not add much latency to the system response. To address these issues, a simple cluster model interpolation based approach for on-line adaptation is presented in this work. The proposed approach employs an OMP based search scheme to select a set of acoustically close models from a set of pre-trained cluster models. The selected cluster models are then linearly interpolated to derive the adapted model parameters. In this work, these interpolation weights are derived from the sparse coefficients in an approximate manner. Such an approximate approach helps in avoiding the iterative ML weight estimation usually employed in existing techniques. The proposed adaptation approach though not optimal, is found to be effective for on-line adaptation. The same has been verified in this work for an LVCSR task and also for an Assamese name recognition system which is a typical example of such query systems.