Performance of automatic Speech Recognition (ASR) systems degrades in noisy conditions due to mismatch between training and test environments. Many methods have been proposed for reducing this mismatch in ASR systems. In recent years, deep neural networks (DNNs) have been widely used in ASR systems and also Robust Speech Recognition and feature extraction. In this paper, we propose to use deep belief network (DBN) as a post-processing method for de-noising Mel frequency cepstral coefficients (MFCCs). In addition, we use deep belief network for extracting tandem features (posterior probability of phones occurrence) from de-noised MFCCs (obtained from previous stage) to obtain more Robust and discriminative features. The final Robust feature vector consists of de-noised MFCCs concatenated to mentioned tandem features. Evaluation results on Aurora2 database show that the proposed feature vector performs better than similar and conventional techniques, where it increases Recognition accuracy in average by 28% in comparison to MFCCs.

The Mel Frequency cepstral coefficients are the most widely used feature in Speech Recognition but they are very sensitive to noise. In this paper to achieve a satisfactorily performance in Automatic Speech Recognition (ASR) applications we introduce a noise Robust new set of MFCC vector estimated through following steps. First, spectral mean normalization is a pre-processing which applies to the noisy original Speech signal. The pre-emphasized original Speech segmented into overlapping time frames, then it is windowed by a modified hamming window .Higher order autocorrelation coefficients are extracted. The next step is to eliminate the lower order of the autocorrelation coefficients. The consequence pass from FFT block and then power spectrum of output is calculated. A Gaussian shape filter bank is applied to the results. Logarithm and two compensator blocks form which one is mean subtraction and the other one are root block applied to the results and DCT transformation is the last step. We use MLP neural network to evaluate the performance of proposed MFCC method and to classify the results. Some Speech Recognition experiments for various tasks indicate that the proposed algorithm is more Robust than traditional ones in noisy condition.

In this work, in order to increase the capacity of a recurrent neural network, we present a model for extracting common features and sharing them across data. As a result of using this model, extracted principle components of data will be invariant to unwanted variations. The recurrent connection of the network removes the noise using a continuous attractor formed during the training phase. The defined speaker codes will be transformed to the information need for switching the continuous attractor in the input space. As a result, speaker variations can be compensated and the Recognition will performed when a clean signal is available. We compared the performance of this method with a reference network described in the paper. The results show that the proposed model is more useful in removing noise and unwanted variations.We compared the performance of this method with the reference network. The results show that the proposed model performs better in removing noise and unwanted variations, it increased the phoneme Recognition accuracy about 5% when the signal to noise ratio is 0 dB.

The conventional view on the problem of Robustness in Speech Recognition is that performance degradation in ASR systems is due to mismatch between training and test conditions. If problem of Robustness in ASR systems is considered as a mismatch between the training and testing conditions the solution would be to find a way to reduce it. Common approaches are: Data-Driven methods such as Speech signal enhancement and using Robust features and model-based methods that alter or adapt model of Speech signal. In this paper, we study a model of environment and obtain a relation between noisy and clean Speech features based on this model. We propose two techniques for mapping noisy features to clean features in cepstrum domain. We implement the proposed methods and some of precedent data-driven methods such as: spectral subtraction, cepstral mean normalization, cepstral mean and variance mean normalization and SNR-dependent cepstral normalization .We show that proposed methods outperform precedent methods and are effective for Robust Speech Recognition in noisy environments.  

Temporal Pattern feature of a Speech signal could be either extracted from the time domain or via their front-end vectors. This feature includes long-term information of variations in the connected Speech units. In this paper, the second approach is followed, i.e. the features which are the cases of temporal computations, consisting of Spectral-based (LFBE) and Cepstrum-based (MFCC) feature vectors, are considered. To extract these features, we use posterior probability-based output of the proposed MTMLP neural networks. The combination of the temporal patterns, which represents the long-term dynamics of the Speech signal, together with some traditional features, composed of the MFCC and its first and second derivatives are evaluated in an ASR task. It is shown that the use of such a combined feature vector results in the increase of the phoneme Recognition accuracy by more than 1 percent regarding the results of the baseline system, which does not benefit from the long-term temporal patterns. In addition, it is shown that the use of extracted features by the proposed method gives Robust Recognition under different noise conditions (by 13 percent) and, therefore, the proposed method is a Robust feature extraction method.