Local linear transformation for voice conversion

Speaker Recognition Anti-spoofing

“…It achieves smooth feature transformations using a local linear transformation. Despite its popularity, known problems of JD-GMM include over-smoothing [73][74][75] and over-fitting [76,77] which has led to the development of alternative linear conversion methods such as partial least square (PLS) regression [76], tensor representation [78], a trajectory hidden Markov model [79], a mixture of factor analysers [80], local linear transformation [73] and a noisy channel model [81]. Non-linear approaches, including artificial neural networks [82,83], support vector regression [84], kernel partial least square [85] and conditional restricted Boltzmann machines [86], have also been studied.…”

Local partial least square regression for spectral mapping in voice conversion

“…One problem of this method is that when the number of nearest neighbourhood is small, over-fitting problem will still be observed. When K is too large, however, over-smoothing problem will occur [9]. Therefore, choosing the optimal number of K is an important issue in local linear transformation method.…”

“…However, over-smoothing and over-fitting problems of JD-GMM method have been reported in many studies [13], [7], [9], [11]. To address the over-fitting problem caused by the full covariance estimation, in [7], partial least square regression method is combined with Gaussian mixture model to replace the transformation matrix estimated by the full covariance matrix, while keeping the mean vectors of the original JD-GMM vector in conversion function.…”

“…This combination, however, cannot avoid the over-smoothing caused by the statistical average during estimating the mean vectors of JD-GMM. One of the successful methods for reducing over-smoothing problem is the local linear transformation method [9]. In this method, each frame has its own transformation matrix, which is estimated from its K-nearest neighbourhood frame pairs in terms of Euclidean distance in the training data.…”

“…A linear conversion function has been implemented by joint density Gaussian mixture model (JD-GMM) with both minimum mean square error and maximum likelihood criteria [1], [2], partial least square regression [7], mixture of factor analyzers [8], local linear transformation [9] and so on. In addition to the linear conversion function, by assuming that the source and target speech features have non-linear relationship, methods such as artificial neural network [10], kernel partial least square [11], and conditional restricted Boltzmann machine [12], have also been proposed.…”