Hardware Speech Sensor Based on Deep Neural Network Feature Extractor and Template Matching

摘要

We explore the possibility of combination of a DNN-based feature extractor and template based matching for keyword detection and speaker recognition embedded in a low power consumption speech sensor. For this task, we can train DNN using a high-performance computer using a large amount of data as we want. However, once it is integrated in the speech sensor hardware, the detection must be performed on an extremely resource-limited condition, and the re-training of the network is prohibited. Therefore, we investigate the use of fixed pre-trained feature extraction DNN and a template-based pattern matching. We evaluate this approach using the CSJ database and FPGA-based implementation.