Transformation of formants for voice conversion using artificial neural networks

Abstract

In this paper we propose a scheme for developing a voice conversion system that converts the speech signal uttered by a source speaker to a speech signal having the voice characteristics of the target speaker. In particular, we address the issue of transformation of the vocal tract system features from one speaker to another. Formants are used to represent the vocal tract system features and a formant vocoder is used for synthesis. The scheme consists of a formant analysis phase, followed by a learning phase in which the implicit formant transformation is captured by a neural network. The transformed formants together with the pitch contour modified to suit the average pitch of the target speaker are used to synthesize speech with the desired vocal tract system characteristics. Zusammenfassung

Voice conversion using Artificial Neural Networks

“…The ANN is trained to map a sequence of source speaker's MCEP's to the target speaker's MCEP's. A generalized back propagation learning law [5] is used to adjust the weights of the neural network so as to minimize the mean squared error between the desired and the actual output values. Selecting initial weights, architecture of the network, learning rate, momentum and number of iterations play an important role in training an ANN [15] .…”

“…But the work presented by [3] proved the need for transformation of excitation features to attain an effective voice morphing system. A variety of techniques have been proposed by researchers for the conversion function, such as mapping code books [1], artificial neural networks [4] [5], dynamic frequency warping [2] or Gaussian mixture model [3] [6] [7] [8].…”

“…Vocal tract shape between two speakers is non linear and hence Artificial Neural Networks (ANN) based method was proposed as they can perform non-linear mapping [5]. Narendranath et.…”

“…al. [5] used ANNs to transform the source speaker formants to target speaker formants. Results were provided showing that the formant contour of the target speaker can be obtained using ANN.…”

“…The use of radial basis function neural network for voice transformation was proposed in [4] [11]. However, the techniques in [5] and [4] used a carefully prepared training data which involved manual selection of vowels or syllable regions from both the source and the target speaker. This is a tedious task to make sure that the source and the target features are aligned correctly.…”