Leveraging gain normalization for sub-band temporal features in noise-robust speech recognition

Accurate steering vector estimation is the key point for a beamformer which suppresses the background noise to improve the noisy speech quality and intelligibility. Recently, time-frequency masking approach, which estimates the steering vectors that are utilized for a beamformer, is popular in this field. In particular, we have proposed an iterative mask estimation (IME) approach to improve the complex Gaussian mixture model (CGMM) based beamforming and yield the best system for multi-channel ASR in CHiME-4 challenge [1]. And in [2], we also demonstrated that our algorithm could improve the speech quality (PESQ) and intelligibility (STOI) for multi-channel speech enhancement. In this study, we focus on the online processing of our IME algorithm for multi-channel speech enhancement and ASR, which achieves comparable performance to the offline version. In addition, a regression long short-term memory recurrent neural network (LSTM-RNN) for a multiple-target joint learning is utilized, denoted as LSTM-MT, to replace two separate models in [2]. Experiments on the CHiME-4 simulation data show that the online IME algorithm can improve the enhancement performance, e.g., PESQ from 2.18 to 2.58 and STOI from 86.85 to 94.51, which is comparable to those obtained by offline IME. Furthermore, the LSTM-MT based post-processing can achieve an additional PESQ improvement from 2.58 to 2.71. And experiments on the CHiME-4 real data show that the online IME approach outperforms the online CGMM-based approach, with a relative word error reduction (WER) of 14.49%.

We propose a novel exemplar-based feature enhancement method for automatic speech recognition which uses coupled dictionaries: an input dictionary containing atoms sampled in the modulation (envelope) spectrogram domain and an output dictionary with atoms in the Mel or full-resolution frequency domain. The input modulation representation is chosen for its separation properties of speech and noise and for its relation with human auditory processing. The output representation is one which can be processed by the ASR back-end. The proposed method was investigated on the AURORA-2 and AURORA-4 databases and improved word error rates (WER) were obtained when compared to the system which uses Mel features in the input exemplars. The paper also proposes a hybrid system which combines the baseline and the proposed algorithm on the AURORA-2 database which in turn also yielded improvement over both the algorithms.

Face recognition in biometrics is a difficult process because of factors including individual differences in facial emotions, illumination, resolution, facial rotation, and facial structure similarities. "A Combination of Cross Domain Features for Face Recognition" is what we suggest here. The benchmarked face images are considered, downsized to consistent 128x128 dimensions, and converted to grayscale. Face photos are compressed, and their quality is improved by using the Discrete Wavelet Transform (DWT). Initial characteristics that are unaffected by changes in an image's lighting are extracted using the Histogram of Oriented Gradients (HOG) on a compressed LL band of DWT. By connecting the LL band's output to the HOG's input, the DWT and HOG's LL bands are connected in a cascade. HOG's input and output are combined to create new, strong final features. The system's results when comparing test and database features are calculated using the Euclidean Distance (ED). It has been noted that the suggested system outperforms the current ones in terms of results.

This paper proposes a new method for feature extraction and recognition of epileptiform activity in EEG signals. The method improves feature extraction speed of epileptiform activity without reducing recognition rate. Firstly, Principal component analysis (PCA) is applied to the original EEG for dimension reduction and to the decorrelation of epileptic EEG and normal EEG. Then discrete wavelet transform (DWT) combined with approximate entropy (ApEn) is performed on epileptic EEG and normal EEG, respectively. At last, Neyman-Pearson criteria are applied to classify epileptic EEG and normal ones. The main procedure is that the principle component of EEG after PCA is decomposed into several sub-band signals using DWT, and ApEn algorithm is applied to the sub-band signals at different wavelet scales. Distinct difference is found between the ApEn values of epileptic and normal EEG. The method allows recognition of epileptiform activities and discriminates them from the normal EEG. The algorithm performs well at epileptiform activity recognition in the clinic EEG data and offers a flexible tool that is intended to be generalized to the simultaneous recognition of many waveforms in EEG.

The histogram equalization approach is an efficient feature normalization technique for noise robust automatic speech recognition. However, it suffers from performance degradation when some fundamental conditions are not satisfied in the test environment. To remedy these limitations of the original histogram equalization methods, class-based histogram equalization approach has been proposed. Although this approach showed substantial performance improvement under noise environments, it still suffers from performance degradation due to the overfitting problem when test data are insufficient. To address this issue, the proposed histogram equalization technique employs the Bayesian estimation method in the test cumulative distribution function estimation. It was reported in a previous study conducted on the Aurora-4 task that the proposed approach provided substantial performance gains in speech recognition systems based on the acoustic modeling of the Gaussian mixture model-hidden Markov model. In this work, the proposed approach was examined in speech recognition systems with deep neural network-hidden Markov model (DNN-HMM), the current mainstream speech recognition approach where it also showed meaningful performance improvement over the conventional maximum likelihood estimation-based method. The fusion of the proposed features with the mel-frequency cepstral coefficients provided additional performance gains in DNN-HMM systems, which otherwise suffer from performance degradation in the clean test condition.

Artificial neural network (ANN) models combined with time series decomposition are widely employed to calculate the river flows; however, the influence of the application of diverse decomposing approaches on assessing correctness is inadequately compared and examined. This study investigates the certainty of monthly streamflow by applying ANNs including feed forward back propagation neural network and radial basis function neural network (RBFNN) models integrated with discrete wavelet transform (DWT), at Jinsha River basin in the upper reaches of Yangtze River of China. The effect of the noise factor of the decomposed time series on the prediction correctness has also been argued in this paper. Data have been analyzed by comparing the simulation outputs of the models with the correlation coefficient (R) root mean square errors, mean absolute errors, mean absolute percentage error and Nash–Sutcliffe Efficiency. Results show that time series decomposition technique DWT contributes in improving the accuracy of streamflow prediction, as compared to single ANN’s. The detailed comparative analysis showed that the RBFNN integrated with DWT has better forecasting capabilities as compared to other developed models. Moreover, for high-precision streamflow prediction, the high-frequency section of the original time series is very crucial, which is understandable in flood season.

This paper proposes a new technique of using discrete wavelet transform (DWT) and back-propagation neural network (BPNN) based on Clarke’s transformation for fault classification and detection on a single circuit transmission line. Simulation and training process for the neural network are done by using PSCAD / EMTDC and MATLAB. Daubechies4 mother wavelet (DB4) is used to decompose the high frequency components of these signals. The wavelet transform coefficients (WTC) and wavelet energy coefficients (WEC) for classification fault and detect patterns used as input for neural network training back-propagation (BPNN). This information is then fed into a neural network to classify the fault condition. A DWT with quasi optimal performance for preprocessing stage are presented. This study also includes a comparison of the results of training BPPN and DWT with and without Clarke’s transformation, where the results show that using Clarke transformation in training will give in a smaller mean square error (MSE) and mean absolute error (MAE). The simulation also shows that the new algorithm is more reliable and accurate.

In this paper, alternative dynamic features for speech recognition are proposed. The goal of this work is to improve speech recognition accuracy by deriving the representation of distinctive dynamic characteristics from a speech spectrum. This work was inspired by two temporal dynamics of a speech signal. One is the highly non-stationary nature of speech, and the other is the inter-frame change of a speech spectrum. We adopt the use of a sub-frame spectrum analyzer to capture very rapid spectral changes within a speech analysis frame. In addition, we attempt to measure spectral fluctuations of a more complex manner as opposed to traditional dynamic features such as delta or double-delta. To evaluate the proposed features, speech recognition tests over smartphone environments were conducted. The experimental results show that the feature streams simply combined with the proposed features are effective for an improvement in the recognition accuracy of a hidden Markov model–based speech recognizer.

Stochastic vector mapping-based feature enhancement using prior-models and model adaptation for noisy speech recognition

This paper investigates the robustness of cepstral based features with respect to additive noise, and details two methods of increasing the robustness with minimal need for a-priori knowledge of the noise statistics. The first approach is a form of noise masking which adds a fixed offset to the linear spectral estimate. The second is a form of sub-band filtering, again in the linear domain, which estimates the dynamic content of the speech using Fourier transforms. This avoids negative values normally inherent in such filtering and which presents difficulties in deriving log estimates. Both methods are shown to provide useful levels of robustness to additive noise, for example, speaker identification error rates in SNR mis-matched conditions of 15 dB are reduced from 60.5% for standard mel cepstra to 13.8% and 24.1% for the two approaches respectively, a relative reduction in error of 77% and 60.1%.

The performance degradation of a speech recognizer in the presence of additive noise is one of the major problems that still remain unsolved in the application of speech recognition technology. This thesis develops speech enhancement schemes where a noisy speech signal is processed in the feature extraction stage. Since the most popular speech features for speech recognition are Mel-Frequency Cepstral Coefficients (MFCC), vectors of Mel-scaled log-spectrum coefficients or cepstrum coefficients are enhanced. Three different speech feature enhancement schemes based on switching linear dynamic models (SLDMs) are proposed. The switching linear dynamic model describes the nonlinear and non-stationary time trajectory of speech features by switching among a set of linear dynamic models over time. With the resulting SLDMs as a speech model and a model for noise, speech and noise can be tracked jointly by means of switching Kalman filtering, which involves a weighted sum of filters operating interactively in parallel. Since the distortion caused by additive ambient noises is highly non-linear in the feature domain, the Extended Kalman Filter algorithm (EKF) and the Unscented Kalman Filter algorithm (UKF) have been used to deal with the nonlinear distortion caused by noise in the feature domain. Comprehensive experiments have been carried out to evaluate the proposed schemes with commonly used databases. The simulation results are presented and compared with other model-based feature enhancement systems in the literature in terms of speech recognition accuracy. Compared with the best results based on the Aurora2 database that we could find in the literature, our approach offers a similar performance (85.96% vs. 86.72%) when the EKF is used in our proposed method, resulting in a smaller complexity. When the UKF is used with our method, our approach achieves a better performance (89.60% vs. 86.72%).

Speech applications in noisy and degraded channel conditions continue to be a challenging problem especially when there is a mismatch between the training and test conditions. In this paper, a robust speech feature extraction scheme is developed based on autoregressive moving average (ARMA) modeling that emphasizes high energy regions of the signal with a data driven modulation filter. The peak preserving ability of two dimensional autoregressive (AR) models is used to emphasize the high energy regions in the spectrotemporal domain. The modulation filtering property is achieved by moving average (MA) modeling. The ARMA spectrograms are used to derive features for speech recognition in the Aurora-4 database. In these experiments, the ARMA model features provide significant improvements (relative improvements of 15%) compared to other robust features. Furthermore, the robustness of these features is also verified for language identification (LID) of highly degraded radio channel speech. Here, the ARMA approach achieves relative improvements of up to 20% over the baseline features.

The electromyography (EMG) signals play a very important role in the muscular activity of the human body. Muscular activities can be measured and analyzed through various movements such as writing, walking, running, hand movements. This paper is focused on hand movements of the amputee people (people with prosthetic hands). Classification of prosthetic hand movements is a research problem that can be solved by classifying the EMG signal generated during movements. The EMG signal has many redundant and irrelevant features that need to be removed. Feature selection (FS) is a technique that is used to select the important features. Therefore, this paper proposes beta artificial bee colony (BetaABC) and binary BetaABC (BBABC) to select prominent features in EMG pattern recognition to increase classification performance and to reduce classifier complexity. BABC is the enhanced version of the ABC inspired from Gaussian ABC (GABC). Initially, discrete wavelet transform (DWT) was applied to decompose and extract the features from EMG signal. After that, BBABC is applied to select the optimal features from the original feature set. The NinaPro dataset (DB3) is used for the experiment and is divided into two sections: (i) Comparative results of proposed BetaABC with ABC and GABC on 10 benchmark functions; (ii) comparative results of proposed BBABC-based feature selection with binary ABC (BABC) and binary GABC (BGABC) on 11 amputee subjects. The obtained results prove that the proposed method is superior to ABC and GABC.

The most common approach for fingerprint analysis is using minutiae that identifies corresponding features and evaluates the resemblance between two fingerprint impressions. Although many minutiae point pattern matching algorithms have been proposed, reliable automatic fingerprint verification remains as a challenging problem. Finger print recognition can be done effectively using texture classification approach. Important aspect here is appropriate selection of features that recognize the finger print. We propose an effective combination of features for multi-scale and multi-directional recognition of fingerprints. The features include standard deviation, kurtosis, and skewness . We apply the method by analyzing the finger prints with discrete wavelet transform (DWT) . We used Canberra distance metric for similarity comparison between the texture classes. We trained 30 images and obtained an overall performance up to 96%.

Invariant feature extraction under diverse illuminations is challenging for face recognition. Related face recognition techniques consider that illumination effect is predominant in low frequencies and involve various methods to segregate high frequency information. However, high frequency feature extraction results in loss of salient features that degrades performance. Thus, objective of this work is to extract illumination normalized robust facial features for face recognition under high illumination conditions. First, a new illumination normalization framework is proposed in which homomorphic filtering (HF) is applied for reducing illumination effect along with contrast enhancement and intensity range compression in face images. Then, illumination deviations are annulled by using reflectance ratio (RR), which yields appropriate texture smoothing and edge preservation. Further, selective feature extraction by discrete wavelet transform (DWT) is performed on HF and RR based face images that discards noise effect. It outcomes in illumination normalized significant facial features, on which subspace analysis (Principal component analysis) is performed to generate small size feature vectors for classification (k-nearest neighbour classifier). Experimental results on benchmark databases such as CMU-PIE, Yale B and Extended Yale B database, demonstrates that proposed face recognition technique yields high performance under diverse illuminations as compared to existing techniques.