Scholarlink

In recent years, the speech recognition also known as Automatic Speech Recognition (ASR) has evolved as a cutting-edge technology. Through the advances of Artificial Intelligence (AI), Deep Learning (DL), and Natural Language Processing (NLP) enhanced speech recognition systems can be developed. The enhanced systems accurately identify phrases, emotions, and sentences. Despite of its advances, speech recognition faces many challenges such as background noise, interferences, language barriers, emotion, and tone recognition. To overcome these problems a Deep Convolutional Neural Network-based Long Short-Term Memory (DCNN-LSTM) is proposed for speech recognition. Initially, a speech recognition (SR) dataset is considered as an input, which consists of audio collected from two distinct datasets. Next, DCNN extracts audio features. Then, the LSTM component is then added to the DCNN acoustic; to learn temporal relations between audio features between audio features, Phonetic Language Modeling (PLM) improves speech recognition accuracy by handling out-of-vocabulary words and captures phonetic patterns and it modeling relationships between words. The proposed DCNN-LSTM model attained better performance metrics of Words Error Rate (WER) of 13.87% when compared with existing method of End-to-End-based Mozilla Deep Speech (E2E-MDS).

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

Speech Recognition Using Deep Convolutional Neural Network-Based Long Short-Term Memory

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Vijaya Bhaskar Reddy Muvva,
  • V. Bala Dhandayuthapani,
  • Sheeba Viswanathan

摘要

In recent years, the speech recognition also known as Automatic Speech Recognition (ASR) has evolved as a cutting-edge technology. Through the advances of Artificial Intelligence (AI), Deep Learning (DL), and Natural Language Processing (NLP) enhanced speech recognition systems can be developed. The enhanced systems accurately identify phrases, emotions, and sentences. Despite of its advances, speech recognition faces many challenges such as background noise, interferences, language barriers, emotion, and tone recognition. To overcome these problems a Deep Convolutional Neural Network-based Long Short-Term Memory (DCNN-LSTM) is proposed for speech recognition. Initially, a speech recognition (SR) dataset is considered as an input, which consists of audio collected from two distinct datasets. Next, DCNN extracts audio features. Then, the LSTM component is then added to the DCNN acoustic; to learn temporal relations between audio features between audio features, Phonetic Language Modeling (PLM) improves speech recognition accuracy by handling out-of-vocabulary words and captures phonetic patterns and it modeling relationships between words. The proposed DCNN-LSTM model attained better performance metrics of Words Error Rate (WER) of 13.87% when compared with existing method of End-to-End-based Mozilla Deep Speech (E2E-MDS).