Srinivas Yarramalle

Bio: Srinivas Yarramalle is an academic researcher. The author has contributed to research in topics: Feature selection & Feature extraction. The author has an hindex of 3, co-authored 5 publications receiving 34 citations.

Speaker Emotion Recognition based on Speech Features and Classification Techniques

Abstract: Speech Processing has been developed as one of the vital provision region of Digital Signal Processing. Speaker recognition is the methodology of immediately distinguishing who is talking dependent upon special aspects held in discourse waves. This strategy makes it conceivable to utilize the speaker's voice to check their character and control access to administrations, for example voice dialing, data administrations, voice send, and security control for secret information. A review on speaker recognition and emotion recognition is performed based on past ten years of research work. So far iari is done on text independent and dependent speaker recognition. There are many prosodic features of speech signal that depict the emotion of a speaker. A detailed study on these issues is presented in this paper. Index Terms—Emotion recognition, feature extraction, speaker recognition.

Unsupervised Image Segmentation Method based on Finite Generalized Gaussian Distribution with EM & K-Means Algorithm

Abstract: Summary In Image Processing Model Based Image Segmentation plays a dominant role in Image Analysis and Image Retrieval . Recently much work has been reported regarding Image Segmentation based on Finite Gaussian Mixture Models using EM algorithm. (Yiming Wu et al (2003)) , (Yamazaki.T (1998)). However, in some images the pixel intensities inside the image regions may not be Meso- Kurtic or Bell Shaped, because of the NonGaussian nature . Hence there are some situations where Image Segmentation is to be done with a more Generalized Finite Mixture Distribution. In this article we develop and analyze an image segmentation method based on Finite Generalized Gaussian Mixture Model using EM and K-Means algorithm. The K-Means algorithm is utilized to obtain the number of regions and the initial estimates of the model parameters. The update equations of the model parameters are obtained by using the EM algorithm. The segmentation of the pixels in the image is done by maximizing the component likelihood function. The performance of this method is evaluated through real time data on 3 images by calculating misclassification rate and image quality metrics. It is observed that the proposed method performs much superior to the earlier image segmentation methods.

A Novel Approach for Effective Emotion Recognition Using Double Truncated Gaussian Mixture Model and EEG

Abstract: Most of the models projected in the literature on Emotion Recognition aims at recognizing the emotions from the mobilized persons in noise free environment and is subjected to the emotion recognition of an individual using a single word for testing and training. Literature available to identify the emotions in case of immobilized persons is confined to the results available from the machines only. In this process braincomputer interaction is utilized using neuro-scan machines like Encephalography (EEG), to identify the emotions of immobilized individuals. It uses the physiological signals available from EEG data extracted from the brain signals of immobilized persons and tries to determine the emotions, but these results vary from machine to machine, and there exists no standardization process which can identify the feelings of the brain diseased persons accurately. In this paper a novel method is proposed, Doubly Truncated Gaussian Mixture Model (DT-GMM) to have a complete emotion recognition system which can identify emotions exactly in a noisy environment from both the healthy individuals and sick persons. The results of the proposed system surpassed the accuracy rates of traditional systems.

Optimization of feature subset using HABC for automatic speaker verification

Abstract: Automatic Speaker Verification is the authentication of a claimed identity based on characteristics of voice. A speaker verification system compares a person's voice with a speaker model or stored voiceprint captured during enrollment as well as an imposter model of different voices, genders and phone types. The system then assigns a confidence score and then makes a decision whether to let the person proceed, to ask for additional voice samples or to refuse entry. Feature subset selection is one of the most concerned processes in the overall classification process of a particular problem. It was also named as dimensionality reduction, attribute subset selection and variable subset selection. For automatic speaker verification (ASV), feature subset selection is one of the first modules. The objective of the paper is to select a most relevant subset of features for error-free optimized classification in the speech domain. In this paper a novel method for speaker verification is proposed using Hybrid Ant Bee Colony optimization to increase the verification rate. Equal Error Rate (EER) is the standard measure which evaluates the projected procedure. Speaker verification system's accuracy rates surpassed the results of traditional systems after applying proposed optimization algorithm; the optimized feature subset was 85% with an average accuracy rate of 95.27%.

Multi Objective Optimization Problem resolution based on Hybrid Ant-Bee Colony for Text Independent Speaker Verification

Abstract: Today major section of automatic speaker verification (ASV) research is focused on multiple objectives like optimization of feature subset and minimization of Equal Error Rate (EER). As such, numerous systems for feature dimension reduction are proposed. This includes framework coaching and testing analysis for every feature set that could be a time esurient trip. Because of its significance, the issue of feature selection has been researched by numerous scientists. In this paper, a new feature subset selection procedure is presented. Hybrid of Ant Colony and Artificial Bee Colony optimized the feature subset over 85% thereby decreased the computational complexity of ASV. Additionally an external record is maintained to store non-dominated solution vectors for which concept of Pareto dominance is used. An overall optimization of 87% is achieved thereby improved the recognition rate of ASV.

Automatic Image Segmentation using Wavelets

Abstract: Model-Based image segmentation plays a dominant role in image analysis and image retrieval. To analyze the features of the image, model based segmentation algorithm will be more efficient compared to non-parametric methods. In this paper, we proposed Automatic Image Segmentation using Wavelets (AISWT) to make segmentation fast and simpler. The approximation band of image Discrete Wavelet Transform is considered for segmentation which contains significant information of the input image. The Histogram based algorithm is used to obtain the number of regions and the initial parameters like mean, variance and mixing factor. The final parameters are obtained by using the Expectation and Maximization algorithm. The segmentation of the approximation coefficients is determined by Maximum Likelihood function. It is observed that the proposed method is computationally efficient allowing the segmentation of large images and performs much superior to the earlier image segmentation methods.

Evaluating Gammatone Frequency Cepstral Coefficients with Neural Networks for Emotion Recognition from Speech.

Abstract: Current approaches to speech emotion recognition focus on speech features that can capture the emotional content of a speech signal. Mel Frequency Cepstral Coefficients (MFCCs) are one of the most commonly used representations for audio speech recognition and classification. This paper proposes Gammatone Frequency Cepstral Coefficients (GFCCs) as a potentially better representation of speech signals for emotion recognition. The effectiveness of MFCC and GFCC representations are compared and evaluated over emotion and intensity classification tasks with fully connected and recurrent neural network architectures. The results provide evidence that GFCCs outperform MFCCs in speech emotion recognition.