Scholarlink

<p>Monaural speech enhancement concentrates on improving the quality and Speech Intelligibility (SI) of a single audio channel. With the advancements in telecommunications and voice-controlled systems, this field has gained popularity. Numerous Deep Learning (DL) strategies are proposed to attain high-quality applications like hearing aids, Automated Speech Recognition (ASR) systems, etc. However, speech distortions and noises created during the enhancement process limit the performance. Addressing these challenges, the present research focuses on removing noise interference, reverberation effects, and reconstruction artifacts, as they are usually detrimental to enhancement systems based on a single channel. Intending to address the issues, this paper outlines an advanced DL network for monaural speech enhancement. This study aims to develop a strong and efficient speech enhancement framework that focuses on enhancing speech quality through a combination of innovative data preprocessing, multi-feature extraction, and a sophisticated DL architecture named ResGANet++. The proposed model is built with two basic architectures: (Miles et al. <CitationRef CitationID="CR1">2022</CitationRef>) GAN for data augmentation and ResNet + + for speech enhancement. Besides, this model employs the hybridized Locust-Kite Optimization Algorithm (LKOA) to optimize the hyperparameters. Extensive testing demonstrated that ResGANet + + outperformed existing DL models in PESQ, STOI, MOS, and regression-based error metrics; thus, signifying its effectiveness in enhancing speech quality and intelligibility in monaural speech enhancement tasks.</p>

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

Resganet + + with Locust-Kite Optimization Aalgorithm (LKOA): an enhanced model based on deep learning for monaural speech enhancement

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • J. Sofia Bobby,
  • V. Mythily,
  • S. Purnima,
  • C. L. Annapoorani,
  • V. Subha Ramya,
  • B. Nivetha

摘要

Monaural speech enhancement concentrates on improving the quality and Speech Intelligibility (SI) of a single audio channel. With the advancements in telecommunications and voice-controlled systems, this field has gained popularity. Numerous Deep Learning (DL) strategies are proposed to attain high-quality applications like hearing aids, Automated Speech Recognition (ASR) systems, etc. However, speech distortions and noises created during the enhancement process limit the performance. Addressing these challenges, the present research focuses on removing noise interference, reverberation effects, and reconstruction artifacts, as they are usually detrimental to enhancement systems based on a single channel. Intending to address the issues, this paper outlines an advanced DL network for monaural speech enhancement. This study aims to develop a strong and efficient speech enhancement framework that focuses on enhancing speech quality through a combination of innovative data preprocessing, multi-feature extraction, and a sophisticated DL architecture named ResGANet++. The proposed model is built with two basic architectures: (Miles et al. 2022) GAN for data augmentation and ResNet + + for speech enhancement. Besides, this model employs the hybridized Locust-Kite Optimization Algorithm (LKOA) to optimize the hyperparameters. Extensive testing demonstrated that ResGANet + + outperformed existing DL models in PESQ, STOI, MOS, and regression-based error metrics; thus, signifying its effectiveness in enhancing speech quality and intelligibility in monaural speech enhancement tasks.