A Deep Neural Networks Approach for Speaker Verification on Embedded Devices

This research deals with three challenges for speaker verification (SV): adaptivity, accuracy, and replay attack. We propose a framework consisting of three independent components: wakeword detector, one time password (OTP) block, and speaker identificator. With this architecture, we can customize each component without significant interference to the whole structure. Via these components, the final representation provides a meaningful information about the speaker to help the system verifies accurately. Moreover, the OTP component can generate a random OTP code when the user interacts with the system, so the replay attack technique cannot be used. The experimental result shows that our proposed solution performs well in terms of the accuracy, running time, and temperature of embedded devices.