Model-Based and Data-Driven Techniques for Environment-Robust Automatic Speech Recognition

Abstract

In this thesis, we propose model-based and data-driven techniques for environment-robust automatic speech recognition. The model-based technique is the feature enhancement method in the reverberant noisy environment to improve the performance of Gaussian mixture model-hidden Markov model (HMM) system. It is based on the interacting multiple model (IMM), which was originally developed in single-channel scenario. We extend the single-channel IMM algorithm such that it can handle the multi-channel inputs under the Bayesian framework. The multi-channel IMM algorithm is capable of tracking time-varying room impulse responses and background noises by updating the relevant parameters in an on-line manner. In order to reduce the computation as the number of microphones increases, a computationally efficient algorithm is also devised. In various simulated and real environmental conditions, the performance gain of the proposed method has been confirmed. The data-driven techniques are based on deep neural network (DNN)-HMM hybrid system. In order to enhance the performance of DNN-HMM system in the adverse environments, we propose three techniques. Firstly, we propose a novel supervised pre-training technique for DNN-HMM system to achieve robust speech recognition in adverse environments. In the proposed approach, our aim is to initialize the DNN parameters such that they yield abstract features robust to acoustic environment variations. In order to achieve this, we first derive the abstract features from an early fine-tuned DNN model which is trained based on a clean speech database. By using the derived abstract features as the target values, the standard error back-propagation algorithm with the stochastic gradient descent method is performed to estimate the initial parameters of the DNN. The performance of the proposed algorithm was evaluated on Aurora-4 DB and better results were observed compared to a number of conventional pre-training methods. Secondly, a new DNN-based robust speech recognition approaches taking advantage of noise estimates are proposed. A novel part of the proposed approaches is that the time-varying noise estimates are applied to the DNN as additional inputs. For this, we extract the noise estimates in a frame-by-frame manner from the IMM algorithm which has been known to show good performance in tracking slowly-varying background noise. The performance of the proposed approaches is evaluated on Aurora-4 DB and better performance is observed compared to the conventional DNN-based robust speech recognition algorithms. Finally, a new approach to DNN-based robust speech recognition using soft target labels is proposed. The soft target labeling means that each target value of the DNN output is not restricted to 0 or 1 but takes non negative values in (0,1) and their sum equals 1. In this study, the soft target labels are obtained from the forward-backward algorithm well-known in HMM training. The proposed method makes the DNN training be more robust in noisy and unseen conditions. The performance of the proposed approach was evaluated on Aurora-4 DB and various mismatched noise test conditions, and found better compared to the conventional hard target labeling method. Furthermore, in the data-driven approaches, an integrated technique using above three algorithms and model-based technique is described. In matched and mismatched noise conditions, the performance results are discussed. In matched noise conditions, the initialization method for the DNN was effective to enhance the recognition performance. In mismatched noise conditions, the combination of using the noise estimates as an DNN input and soft target labels showed the best recognition results in all the tested combinations of the proposed techniques.