Conference Papers
Permanent URI for this collectionhttps://idr.nitk.ac.in/handle/123456789/28506
Browse
4 results
Search Results
Item NITK-KLESC: Kannada Language Emotional Speech Corpus for Speaker Recognition(Institute of Electrical and Electronics Engineers Inc., 2023) Tomar, S.; Gupta, P.; Koolagudi, S.G.This work introduces an emotional speech dataset for Speaker Recognition (SR) task. The proposed dataset is recorded in the Kannada language from the people of Karnataka state of India. The speech dataset is collected by simulating five different emotions, such as Fear, Sad, Anger, Happy, and Neutral. The dataset is named as National Institute of Technology Karnataka, India- Kannada Language Emotional Speech Corpus (NITK-KLESC). The proposed dataset will be useful for SR tasks in various emotions. The proposed emotional speech dataset will be useful for emotion recognition, analysis of emotional speech, speech recognition, gender identification, and age identification of the age group 20 to 50 years. The proposed work describes the development, processing, analysis, acquisition, and evaluation of the proposed emotional speech dataset (NITK-KLESC). The analysis of emotional speech was done by considering various basic speech parameters like Pitch, Tempo, Intensity, and Zero Crossing Rate (ZCR). The characteristics of the dataset are reported using MFCC feature extraction and considered the CNN model as a classifier, compared with the existing EmoDB dataset. The average accuracy of the Emotional Speech Speaker Recognition (ESSR) task was measured at 84.44% with the EmoDB dataset and 95.2% with the proposed NITK-KLESC dataset. © 2023 IEEE.Item NITK-TIEKLS: A Text-Independent Emotional Kannada Language Speech Dataset for Speaker Recognition(Springer Science and Business Media Deutschland GmbH, 2025) Tomar, S.; Koolagudi, S.G.Speaker recognition systems have traditionally relied on the consistency of speech content to identify individuals. However, text-independent speaker recognition, irrespective of the spoken content, presents a more flexible and robust alternative, especially in real-world scenarios. This research focuses on enhancing text-independent speaker recognition by incorporating a dataset for the Speaker Recognition (SR) task. The dataset is named the National Institute of Technology Karnataka - Text-Independent Emotional Kannada Language Speech (NITK-TIEKLS) dataset. The 200 natives of the Karnataka state of India have recorded emotional speech in the Kannada language for the proposed dataset. The neutral text-independent speech consists of a 4-min speech duration for each speaker. The two emotional speech utterances, from any two of the emotions anger, happiness, sadness, and fear, are text-independent speech utterances that consist of 2 min. The total duration is approximately 30 h. The proposed study includes developing, processing, analyzing, acquiring, and evaluating the proposed dataset. The suggested dataset consists of performance evaluations of the SR system through deep learning techniques with the proposed Wavelet-Mel Spectrogram. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.Item Identification of Speaker-Specific Features to Minimize the Mismatch Outcomes for Speaker Recognition Using Anger and Happy Emotional Speech(Springer Science and Business Media Deutschland GmbH, 2025) Tomar, S.; Koolagudi, S.G.A vital component of digital speech processing is Speaker Recognition (SR). However, variation in speakers’ emotional states, such as happiness, anger, sadness, or fear, poses a significant challenge that compromises the robustness of speaker recognition systems. It appears to be challenging to distinguish between emotions like “anger†and “happy†, according to research on SR using emotive speech. The study looks at prosody-related speech characteristics to determine how to distinguish between “anger† and “happy†emotional speech for SR tasks. The goal is to explore speaker-specific features. The experiment outcomes demonstrate that, as speaker-specific features for the SR task, Intensity, Pitch, and Brightness (IPB) variables can distinguish between angry and happy emotional speech. Combining IPB and MFCC (IPBCC) feature extraction with the Hybrid CNN-LSTM combined with an attention mechanism approach achieves an SR accuracy of 95.45% for anger and 96.22% for happy emotional speech. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.Item Transformation of Emotional Speech to Anger Speech to Reduce Mismatches in Testing and Enrollment Speech for Speaker Recognition System(Springer Science and Business Media Deutschland GmbH, 2025) Tomar, S.; Koolagudi, S.G.Speaker Recognition (SR) is a critical component of digital speech processing. The robustness of Speaker Recognition systems is compromised by the variance in speakers’ emotional states. According to a study on SR utilizing emotive speech, it seems complicated to distinguish between emotions like “anger,†“sad,†“fear,†and “happy†. Developing a speaker recognition model that works effectively using emotional speech is challenging, specifically in the case of some intense emotions like anger. This work explores emotional speech transformation approaches to reduce the mismatch between training and testing emotional speech for the SR tasks. The recommended effort aims to develop speech transformation techniques to transform different emotional speech into anger. This study modifies the prosodic features “TPIB†(Tempo, Pitch, Intensity, and Brightness) to transform the speech from neutral, happy, fearful, and sad emotions to anger. Performance evaluations of the SR system employing transformed emotional speech are obtained through integrating Mel-Spectrogram feature extraction and deep learning techniques, including the CREMA-D and NITK-KLESC datasets. The experiment results demonstrate that the suggested emotional speech transformation technique increases SR accuracy in transforming neutral by approximately 15%, happy by 11%, sad by 32%, and fear by 30%. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.