Total variational neighborhood optimization in LLE manifold learning for acoustic feature extraction

To address the issue of low efficiency and accuracy in acoustic feature extraction in manifold learning, a novel full variational manifold learning approach is extracted. Based on the local linear embedding manifold, a high-dimensional matrix of acoustic feature manifolds is constructed from short-time Fourier time-frequency data. For this matrix, data low-dimensional mapping is achieved through high-order total variation and total variation regularization manifolds. The mathematical recursive formula of high-order variation is deduced by utilizing the differentiability of one-dimensional variation, extending to form a high-order total variation matrix. The selection range of the neighborhood of the clustering accuracy quantitative manifold is constructed based on the clustering property of total variation. This process converts the neighborhood selection from the direct distance measurement of the traditional <i>k</i>-nearest neighbors (KNN) to the distance gradient measurement to facilitate data clustering. The neighborhood weights of sample data are optimized with the high-order total variation matrix, and the local linear mapping function for dimensionality reduction and reconstruction is constructed to ensure the consistency of local relationships in the reduced-dimensional data. A total variation manifold regularization model is constructed based on the regularization concept, and a manifold is generated through convex optimization. The validity of the method is verified through numerical simulation and experiments.