Machine learning technologies have improved the accuracy of speech recognition systems, and devices using those systems, such as smart speakers and AI assistants, are now in wide use. However, speech recognition systems have security vulnerabilities. In particular, a known machine learning vulnerability called audio adversarial examples (AAEs), which causes misrecognition in speech recognition systems, has become a problem. We propose a scheme for using speech processing to protect speech recognition systems from AAEs, preventing misrecognitions by slight processing of input speech that does not affect the recognition of normal speech. We use two kinds of processing: speed and frequency. Evaluation results show that the proposed scheme can reduce the success rate of attack speech to about 1% while maintaining about 85% recognition rates for normal speech.
