Bio-Inspired Sensor-Based Sound Pre-Processing for Speech Recognition in Noisy Conditions

Despite tremendous research efforts in the field of speech recognition systems, current technologies cannot match the performance of human hearing in noisy environments. The recently introduced adaptive microelectromechanical (MEMS) cochlea [1] offers a potential solution here, as it performs signal (pre-)processing directly by the sensor, inspired by the functionality of the human cochlea. To show these advantages, we investigate here in a combined experimental and theoretical study, the speech recognition performance as a function of pre-processing. For speech recognition, a reservoir computing model based on memristive devices is used. For pre-processing, the MEMS cochlea is compared with a state-of-the-art Mel Frequency Cepstral Coefficients (MFCC) filter. For clean speech the system using MFCC filter has a slightly higher recognition efficiency than the one using the MEMS cochlea. In noisy environments, however, the MEMS cochlea is ahead. Thus, the MEMS cochlea-based pre-processing offers a route to hardware-implemented, edge computing for speech processing at the sensor node with high efficiency and low latency.