Scholarlink

The usage of speech data with the emerging advancement of Artificial Intelligence has significantly increased in diverse sectors, placing speech sentiment analysis as a prominent research focus within the realm of Natural Language Processing (NLP). Though Bangla is spoken by approximately 210 million people worldwide, Bangla speech sentiment analysis is still limited in research. This paper addresses the research gap by leveraging the strengths of Convolutional Neural Networks (CNN) and Transformer architectures, applied to the largest publicly available corpora in this language: SUBESO and BanglaSER. Relying on Mel-Frequency Cepstral Coefficients (MFCC), our model is trained on both datasets. This model has been evaluated on 7 distinct emotions: Surprise, Neutral, Sad, Happy, Angry, Disgust, and Fear. Our model has achieved 92% accuracy on the SUBESCO dataset. In addition, this is the first model to achieve 91% accuracy with the combined dataset (SUBESCO+BanglaSER) on 7-class Bangla speech emotion recognition using a hybrid CNN-Transformer architecture.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging MFCC Features for Bangla Audio Emotion Classification Using a ResNet-Transformer Encoder Fusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Md Rezaul Hossain,
  • Fakiha Rahman Soha,
  • Tajkiratul Abida Ananna,
  • Ruba Sazeda,
  • Anisur Rahman

摘要

The usage of speech data with the emerging advancement of Artificial Intelligence has significantly increased in diverse sectors, placing speech sentiment analysis as a prominent research focus within the realm of Natural Language Processing (NLP). Though Bangla is spoken by approximately 210 million people worldwide, Bangla speech sentiment analysis is still limited in research. This paper addresses the research gap by leveraging the strengths of Convolutional Neural Networks (CNN) and Transformer architectures, applied to the largest publicly available corpora in this language: SUBESO and BanglaSER. Relying on Mel-Frequency Cepstral Coefficients (MFCC), our model is trained on both datasets. This model has been evaluated on 7 distinct emotions: Surprise, Neutral, Sad, Happy, Angry, Disgust, and Fear. Our model has achieved 92% accuracy on the SUBESCO dataset. In addition, this is the first model to achieve 91% accuracy with the combined dataset (SUBESCO+BanglaSER) on 7-class Bangla speech emotion recognition using a hybrid CNN-Transformer architecture.