Front End Audio Processing for Voice Enabled Products

Presenter:
Paul Beckmann, DSP Concepts, Inc. - Santa Clara, CA USA

Voice recognition has become a sought-after feature in consumer and automotive audio products. Many OEMs are now scrambling to add these features to their products with little or no experience with microphone processing and many are struggling. This session focuses on the front end audio processing needed by a device to properly interface to a cloud based ASR engine. We cover beamforming, echo cancellation, direction of arrival estimation, and noise reduction. We show how the algorithms must be designed to work in concert for far field voice pickup and the difficult to achieve "barge in" feature. Performance metrics and evaluation procedures for the various algorithms are presented. Particular emphasis is given to the design of the microphone arrays and beamforming. We also present a novel metric that is correlated with performance and allows easy comparison of beamformer designs.