Speaker clustering for the detection of mild cognitive impairment based on deep neural networks and audio signal
Hong-Han Hank Chau, Yawgeng A. Chau · IET conference proceedings. · 2024
In this paper, we propose speaker identification through speaker clustering for the detection of mild cognitive impairment (MCI). For dataset collection, we cooperated with a neurologist from a local hospital to obtain total 110 voice recordings of potential patients, including 79 MCI cases. We first extract 256 speaker features using the speaker embedding generated from the Pyannote toolkit with transfer learning and tuned parameters. Then, these speaker features are employed for speaker clustering by using K-means methods. To study the performance of speaker identification, the relevant confusion matrix is derived and the accuracy index is given. From the accuracy analysis, we can achieve 95% accuracy for the speaker identification, and 88% accuracy for the MCI detection using the multilayer perceptron (MLP) deep neural network.