Event - Driven Predictive Sampling Circuits for Speech Signal Sensing and Processing

This paper presents a novel event-driven speech signal sensing, pre-processing, and compression method using Dynamic Predictive Sampling. The Dynamic Predictive Sampling method converts the input analog waveform into a non-uniform sampled event sequence with both amplitude and timestamp data of each event, which is the turning point of the analog signal. The event sequence can be reconstructed without losing the morphology of the input analog signal. Since the selection of turning points is performed during the analog-to-digital conversion process, the circuit generates much less data throughput. This paper studies the trade-off between the compression factor and the performance in speech recognition accuracy of the proposed method. Based on the simulation result, the total data throughput can be reduced by 87% while keeping the quality of the speech signal for speech recognition. An integrated circuit of Dynamic Predictive Sampling has been designed and simulated for the speech sensing task. The proposed method saves computing overhead and data throughput, which is ideal for future low-power embedded voice recognition systems.