Machine learning (ML) has opened avenues for performing various prediction/recognitions such as speech recognition, disease diagnosis, and drug prescription without explicit instructions. In the coming era of AIoT, the AI-based sensors will play an innovative platforms via the synergistic collaboration between big-data and real-time processing.
We have demonstrated speaker recognition of HAND’s acoustic sensor with the Gaussian mixture model (GMM) calculating and visualizing the differences among voice features. The similar speech signals are classified into the cluster of subpopulations, which represents the probability distributions. Other ML algorithms could be adopted to enhance the versatility and capability of HAND’s sensors. The wearable spignanometer enables the exact diagnosis on the critical hypertension using support vector machine (SVM) that calculates the decision boundary; the Hidden Markov model (HMM) can allow the accurate speech recognition of flexible piezoelectric acoustic sensor based on the sequential voice data processing.
Recently, deep learning (DL) algorithms have been spotlighted as a core technology of AIoT service platforms since the DL architectures computes complex tasks and further exceed the human performance by using the artificial neural networks. Our group demonstrated the noise-robust speech processing of multi-channel sensor via convolutional neural network (CNN) and deep U-net. Our CNN-assisted acoustic sensor, calculating the optimized weights on each channel, showed the minimal loss of accuracy rate even at harsh noisy condition. We successfully filtered the undesirable background noises from multi-channel speech signals via deep U-net, which indicates the potential for real-life VUI applications of our sensor in social communications, business meeting, and court trials.
[Related References]
"Machine Learning-based Self-powered Acoustic Sensor for Speaker Recognition", Nano Energy, 53, 658, 2018
“Flexible Piezoelectric Acoustic Sensors and Machine Learning for Speech Processing”, Adv. Mater., 32, 1904020, 2020
"Biomimetic and flexible piezoelectric mobile acoustic sensors with multiresonant ultrathin structures for machine learning biometrics", Sci. Adv., 7, eabe5683, 2021
"Deep learning-based noise robust flexible piezoelectric acoustic sensors for speech processing", Nano Energy, 101, 107610, 2022
Machine learning (ML) has opened avenues for performing various prediction/recognitions such as speech recognition, disease diagnosis, and drug prescription without explicit instructions. In the coming era of AIoT, the AI-based sensors will play an innovative platforms via the synergistic collaboration between big-data and real-time processing.
We have demonstrated speaker recognition of HAND’s acoustic sensor with the Gaussian mixture model (GMM) calculating and visualizing the differences among voice features. The similar speech signals are classified into the cluster of subpopulations, which represents the probability distributions. Other ML algorithms could be adopted to enhance the versatility and capability of HAND’s sensors. The wearable spignanometer enables the exact diagnosis on the critical hypertension using support vector machine (SVM) that calculates the decision boundary; the Hidden Markov model (HMM) can allow the accurate speech recognition of flexible piezoelectric acoustic sensor based on the sequential voice data processing.
Recently, deep learning (DL) algorithms have been spotlighted as a core technology of AIoT service platforms since the DL architectures computes complex tasks and further exceed the human performance by using the artificial neural networks. Our group demonstrated the noise-robust speech processing of multi-channel sensor via convolutional neural network (CNN) and deep U-net. Our CNN-assisted acoustic sensor, calculating the optimized weights on each channel, showed the minimal loss of accuracy rate even at harsh noisy condition. We successfully filtered the undesirable background noises from multi-channel speech signals via deep U-net, which indicates the potential for real-life VUI applications of our sensor in social communications, business meeting, and court trials.
[Related References]
"Machine Learning-based Self-powered Acoustic Sensor for Speaker Recognition", Nano Energy, 53, 658, 2018
“Flexible Piezoelectric Acoustic Sensors and Machine Learning for Speech Processing”, Adv. Mater., 32, 1904020, 2020
"Biomimetic and flexible piezoelectric mobile acoustic sensors with multiresonant ultrathin structures for machine learning biometrics", Sci. Adv., 7, eabe5683, 2021
"Deep learning-based noise robust flexible piezoelectric acoustic sensors for speech processing", Nano Energy, 101, 107610, 2022