Machine learning techniques have long been applied in many fields and have gained a lot of success. The purpose of learning processes is generally to obtain a set of parameters based on a given data set by minimizing a certain objective function which can explain the data set in a maximum likelihood or minimum estimation error sense. However, most of the learned parameters are highly data dependent and rarely reflect the true physical mechanism that is involved in the observation data. In order to obtain the inherent knowledge involved in the observed data, it is necessary to combine physical models with learning process rather than only fitting the observations with a black box model. To reveal underlying properties of human speech production, we proposed a learning process based on a physiological articulatory model and a coarticulation model, where both of the models are derived from human mechanisms. A two-layer learning framework was designed to learn the parameters concerned with physiological level using the physiological articulatory model and the parameters in the motor planning level using the coarticulation model. The learning process was carried out on an articulatory database of human speech production. The learned parameters were evaluated by numerical experiments and listening tests. The phonetic targets obtained in the planning stage provided an evidence for understanding the virtual targets of human speech production. As a result, the model based learning process reveals the inherent mechanism of the human speech via the learned parameters with certain physical meaning.

Among various training concepts for speaker adaptation, Speaker Adaptive Training (SAT) has been successfully applied to a standard Hidden Markov Model (HMM) speech recognizer, whose state is associated with Gaussian Mixture Models (GMMs). On the other hand, focusing on the high discriminative power of Deep Neural Networks (DNNs), a new type of speech recognizer structure, which combines DNNs and HMMs, has been vigorously investigated in the speaker adaptation research field. Along these two lines, it is natural to conceive of further improvement to a DNN-HMM recognizer by employing the training concept of SAT. In this paper, we propose a novel speaker adaptation scheme that applies SAT to a DNN-HMM recognizer. Our SAT scheme allocates a Speaker Dependent (SD) module to one of the intermediate layers of DNN, treats its remaining layers as a Speaker Independent (SI) module, and jointly trains the SD and SI modules while switching the SD module in a speaker-by-speaker manner. We implement the scheme using a DNN-HMM recognizer, whose DNN has seven layers, and elaborate its utility over TED Talks corpus data. Our experimental results show that in the supervised adaptation scenario, our Speaker-Adapted (SA) SAT-based recognizer reduces the word error rate of the baseline SI recognizer and the lowest word error rate of the SA SI recognizer by 8.4% and 0.7%, respectively, and by 6.4% and 0.6% in the unsupervised adaptation scenario. The error reductions gained by our SA-SAT-based recognizers proved to be significant by statistical testing. The results also show that our SAT-based adaptation outperforms, regardless of the SD module layer selection, its counterpart SI-based adaptation, and that the inner layers of DNN seem more suitable for SD module allocation than the outer layers.

How to reduce noise with less speech distortion is a challenging issue for speech enhancement. We propose a novel approach for reducing noise with the cost of less speech distortion. A noise signal can generally be considered to consist of two components, a "white-like" component with a uniform energy distribution and a "color" component with a concentrated energy distribution in some frequency bands. An approach based on noise eigenspace projections is proposed to pack the color component into a subspace, named "noise subspace". This subspace is then removed from the eigenspace to reduce the color component. For the white-like component, a conventional enhancement algorithm is adopted as a complementary processor. We tested our algorithm on a speech enhancement task using speech data from the Texas Instruments and Massachusetts Institute of Technology (TIMIT) dataset and noise data from NOISEX-92. The experimental results show that the proposed algorithm efficiently reduces noise with little speech distortion. Objective and subjective evaluations confirmed that the proposed algorithm outperformed conventional enhancement algorithms.