Scholarlink

<p>In the digital era, Speech-To-Text (STT) conversion has emerged as a significant research area with a wide range of applications. However, achieving accurate speech recognition and rapid conversion presents several challenges, including variations in accents, mixed-language usage, and differences in pronunciation, etc. Therefore, the proposed speech-to-text conversion system leverages the use of Deep Learning (DL) Networks for enhancing the accuracy, reducing the processing time and error rates during speech processing. The pre-processing at the first step for AMWPMF filters out unwanted noises, and then acoustic features related to speech such as MFCCs and LPCCs are extracted. Finally, speech-to-text conversion is performed using the help of Shunted Self-Attention (SSA) based Graph Convolutional Neural Network (GCNN), and its parameters are updated with Humboldt Squid Optimization Algorithm (HSOA) for better efficiency with lesser error rates. The efficiency of the proposed speech-to-text conversion system is analyzed in terms of various metrics like error rates, processing time, accuracy, precision, f1-score, and recall with various existing state-of-art speech-to-text conversion systems under MuST-C and LibriSpeech datasets. The results show that the suggested system works better than the compared current systems, offering 96% more accuracy, 94% more precision, 95% f1-score, 97% recall, and 13 seconds less processing time, respectively.</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

Optimized graph convolutional shunted self-attention neural network for multilingual speech-to-text training using cross-language voice conversion of speech representations

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Selvan Chinnaiyan,
  • Anwar Basha Haffishthullah,
  • Soumyalatha Naveen,
  • Mohamed Imtiaz Naseeruddin

摘要

In the digital era, Speech-To-Text (STT) conversion has emerged as a significant research area with a wide range of applications. However, achieving accurate speech recognition and rapid conversion presents several challenges, including variations in accents, mixed-language usage, and differences in pronunciation, etc. Therefore, the proposed speech-to-text conversion system leverages the use of Deep Learning (DL) Networks for enhancing the accuracy, reducing the processing time and error rates during speech processing. The pre-processing at the first step for AMWPMF filters out unwanted noises, and then acoustic features related to speech such as MFCCs and LPCCs are extracted. Finally, speech-to-text conversion is performed using the help of Shunted Self-Attention (SSA) based Graph Convolutional Neural Network (GCNN), and its parameters are updated with Humboldt Squid Optimization Algorithm (HSOA) for better efficiency with lesser error rates. The efficiency of the proposed speech-to-text conversion system is analyzed in terms of various metrics like error rates, processing time, accuracy, precision, f1-score, and recall with various existing state-of-art speech-to-text conversion systems under MuST-C and LibriSpeech datasets. The results show that the suggested system works better than the compared current systems, offering 96% more accuracy, 94% more precision, 95% f1-score, 97% recall, and 13 seconds less processing time, respectively.