Speech sampling by level-crossing and its reconstruction using spline-based filtering

The paper discusses an original approach to speech processing. Traditionally it is organized using uniform sampling with a fixed sampling rate, which is followed by encoding to reduce amount of data for transmission. The proposed approach is based on event-driven digital-to-analog conversion that produces data with low redundancy and allows the compression procedure to be omitted. The application of the level-crossing method provides samples without quantization errors in amplitude. However, since they are spaced non-equidistantly, a dedicated reconstruction algorithm is required. A recovery procedure, which uses spline-based filtering with time-varying bandwidth, has been developed. The instantaneous maximum frequency of a signal is estimated from the local sampling density. The results of the simulation produced using the proposed method are presented. The developed approach can be implemented using asynchronous design techniques, and is targeted at application in speech transmission over wireless networks.