Study And Application Of Formant Tracking Technology

Formant parameters and trajectories are important features of phonetics and have a wide range of applications in the speech signal processing field. Therefore, reliable formant tracking algorithm would be the basis of various studies on speech.Abundant researches of acoustic phonetics and articulatory phonetics show that the production process of speech signals are nonlinear and the background noises are non-Gaussian. In general, the problem of formant tracking can be regarded as a version of dynamic state estimation. Bayesian methods provide a rigorous general framework for dynamic state estimation problems. The particle filter method is to complete the process of Bayesian recursion based on Monte Carlo simulation, which isn't restricted of linearity of the model or Gaussian noises.This dissertation presents methods for formant tracking by using particle filters technology. Firstly, a hidden dynamic model, with state-space equations, is introduced into describing the dynamic characteristics of the speech signals. The framework for formant tracking, based on particle filters, is established by using the formant of the speech signals as the hidden state-variables and the nonlinear mapping from formant to linear prediction cepstral coefficients as the observation equations. Then the formant tracking is performed by using elementary particle filters. The theoretical and experimental analyses indicate that the particles of elementary particle filters are generated from the prior distribution can easily land on low-likelihood areas thus wasted. To overcome this difficulty, a method for formant tracking based on unscented particle filter is proposed. However, the performance of the modified method is not satisfactory in the formant tracking of spontaneous speech. This dissertation further proposes a new algorithm to generate sophisticated proposal distribution that integrates the phoneme information by introducing an auxiliary vector, thus greatly improving the tracking performance. The auxiliary particle filter allows us to obtain approximate samples from the optimal importance distribution by using an auxiliary vector. The joint use of the dynamic formant prior and auxiliary information makes the distribution of the initial particles closer to the true distribution, and accurate experimental results will be achieved by using a small number of particles.Finally, this dissertation proposes an application method of formant parameters and trajectories for voice conversion and Chinese whispers recognition, respectively. In voice conversion system, the formant parameters and pitch are considered as the transformation parameters of the conversion process with which the transfer function is deducted by Gaussian mixture model. By evaluating subjectively and objectively, the transformed voices sound more naturally, the intelligibility and orientation have also been improved. In Chinese whispers recognition system based on Gaussian mixture model, linear prediction cepstral coefficients, Mel-frequency cepstral coefficients and amended Mel-frequency cepstral coefficients are considered as the parameters for the first-class recognition. The second-class recognition is performed after the first-class recognition using the trajectories of formant as the judgement rules. Experimental results show that the rate of Chinese whispers recognition has been improved by using the second-class recognition based on the trajectories of formant.