| 초록 |
Voice recognition is used as most direct bilateral communication method between humans and smart devices. For the applications such as personalized voice-controlled assistant, smart home appliance that based on artificial intelligence (AI) and internet of things (IoT), speaker recognition is essential element. Here in, flexible piezoelectric acoustic sensors (f-PAS) inspired by basilar membrane in the human cochlear are used for speaker recognition by a new concept of machine learning. The f-PAS are resonant-type sensor and self-powered that have high sensitivity. The f-PAS broadly covered the voice frequency spectrum of human with four to eight times higher sensitivity through the combination of multi-resonant frequency tuning and low quality factor (Q). Via multi-channel sound input, f-PAS acquired abundant voice information. The standard TIDIGITs dataset were used for training and testing speaker recognition performance of f-PAS. The TIDIGITs were recorded by the f-PAS in free field condition and changeover to frequency components by Fast Fourier Transform (FFT) and Short-Time Fourier Transform (STFT) methods. Using the average value of most highest and second highest output among multi-channel data, the machine learning algorithm was put in to practice for speaker recognition. The f-PAS based speaker recognition exhibit magnificent recognition rate of 97.5% with error rate 2.5%, which is reduced 75% compared to that of reference condense-type MEMS microphone. |
