Search

Emotion recognition is one of the most challenging tasks in the field of speech recognition and processing. Besides, it is deemed a significant branch of Human-Computer Interaction (HCI). Although there exist several techniques related to speech emotion recognition (SER), in this study, machine learning approaches for feature extraction is used. To be precise, the Mel-frequency Cepstrum Coefficient (MFCC) and Modulation Spectral (MS) is implemented to extract the relevant features from Bengali speech. Our study focuses on recognizing several states of emotion such as joy, sadness, anger, surprise, fear, and disgust from Bengali speech by exploring an emotion recognition model. Besides, in this study, a novel Bengali speech dataset is developed and used since no proper study, specifically for Bengali, on this topic is available in the literature yet. In this study, the recurrent neural network (RNN) technology is used to categorize the human voice (speech) into six particular states of emotion. Finally, our study shows that while the RNN classifier model achieves 47.66% accuracy with our Bengali speech dataset, achieved 51.33% accuracy when the dataset is first categorized using the RNN modulation technique.

Speaking is undeniably an effective way of communication in any language, which is now an essential feature in various sectors of technology. This study proposes to build such a speech to text conversion system for the Bengali language because Bengali, being one of the most popular languages worldwide, is very little explored in this research arena. Hence, the study aims to provide a platform to detect speech and recognize its content as text automatically. This work promises to help the mass people who may not be educated enough to write fluently and accurately. In this study, the focus is on processing Bengali speech data using ‘DeepSpeech’, which creates a neural network to recognize the audio files containing speech and then, to transform the audio speech into its text format. Moreover, the performance evaluation of the proposed system is accomplished after integrating it with the proposed Bengali spell corrector mechanism to achieve significant improvement over the existing approach.

Detecting parts of speech of words in a sentence efficiently is an inseparable task in natural language processing (NLP). The goal of this study is to detect the parts of speech of words in a Bengali sentence with higher accuracy and memory optimization. In this study, we develop such a parts of speech tagging tool specifically for those people who want to learn or research on Bengali grammar. Besides, it has an impact on the quality of education as our tool aims to provide knowledge about Bengali grammar. Although there are different existing studies on Bengali parts of speech tagging such as hybrid model, global linear model, maximum entropy model, hidden Markov model, unsupervised model, Bangla parsing algorithm, rule-based classifier, decision tree algorithm, Baum–Welch algorithm, unsupervised word segmentation algorithm, etc., none of these studies consider memory optimization technique. Therefore, in this study, we propose and endeavor to implement our own rule-based parts of speech tagging mechanism with an intent to achieve higher accuracy by optimizing memory overhead.

Basic human values represent a set of values such as security, independence, success, kindness, and pleasure, which we deem important to our lives. Each of us holds different values with different degrees of significance. Existing studies show that values of a person can be identified from their social network usage. However, the value priority of a person may change over time due to different factors such as life experiences, influence, social structure and technology. Existing studies do not conduct any analysis regarding the change of users' value from the social influence, i.e., group persuasion, form the social media usage. In our research, first, we predict users' value score by the influence of friends from their social media usage. We propose a Bounded Confidence Model (BCM) based value dynamics model from 275 different ego networks in Facebook that predicts how social influence may persuade a person to change their value over time. Then, to predict better, we use particle swarm optimization based hyperparameter tuning technique. We observe that these optimized hyperparameters produce accurate future value score. We also run our approach with different machine learning based methods and find support vector regression (SVR) outperforms other regressor models. By using SVR with the best hyperparameters of BCM model, we find the lowest Mean Squared Error (MSE) score 0.00347., Comment: 11 pages

Video captioning refers to the process of predicting a semantically consistent textual description from a given video clip. Even though a significant amount of research work is present for video captioning in English, for Bengali the field of video captioning is nearly unexplored. Therefore, this research aims at generating Bengali captions that plausibly describe the gist of a specific short video. To accomplish this, Long Short-Term Memory (LSTM) based a sequence-to-sequence model is used that takes the video frame features as input and generates an analogous textual description. In this study, Microsoft Research Video Description Corpus (MSVD) dataset is used which is an English dataset. Therefore, a deep learning-based translator and manual labor are used to convert English captions into appropriate Bengali ones. Finally, the model's performance is evaluated using popular evaluation metrics - BLEU and TER. The proposed approach achieves BLEU and TER scores of 0.38 and 0.76 respectively, establishing a new benchmark for the Bengali video captioning tasks.