Scholarlink

<p>Speech-to-text permits the transcription of voice inputs into text. Specialists in various industries who require precise textual translations utilize this service. This rapidly developing technology enables people with disabilities to communicate in a natural and intuitive manner similar to human speech. This technology can be particularly beneficial for those who face communication challenges. In this work a lightweight end-to-end deep convolutional neural network (DCNN) is suggested to provide flexibility and ease of communication. To map the spoken words into appropriate text transcription an end-to-end DCNN architecture has been modified layer by layer and incorporates a Gated Recurrent Unit (GRU) approach. The suggested bidirectional GRU layer is a computationally effective framework used in lightweight DCNNs to enhance overall system efficiency which also has a significant role in speech-to-text recognition. This work employed the standard CHiME-5 corpus comprises multi-speaker conversational recordings across many real-home dinner parties to assess the proposed framework. The word error rate (WER) of 34.65% was obtained by applying this dataset to the proposed bidirectional GRU-based DCNN framework. Also the suggested model performance is validated using the TED-LIUM andLibriSpeech datasets. Our proposed model achieved 10 .65% WER on the TED-LIUM dataset and 10.08% WER on LibriSpeech, respectively, which is a relative improvement in performance over baseline models.</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

A lightweight bidirectional GRU–DCNN hybrid framework for end-to-end automatic speech recognition

  • 期刊论文

  • 发表日期：

    2026年4月11日

查看全文
  • V. Karthikeyan,
  • P. Saranya,
  • M. Natchiyar

摘要

Speech-to-text permits the transcription of voice inputs into text. Specialists in various industries who require precise textual translations utilize this service. This rapidly developing technology enables people with disabilities to communicate in a natural and intuitive manner similar to human speech. This technology can be particularly beneficial for those who face communication challenges. In this work a lightweight end-to-end deep convolutional neural network (DCNN) is suggested to provide flexibility and ease of communication. To map the spoken words into appropriate text transcription an end-to-end DCNN architecture has been modified layer by layer and incorporates a Gated Recurrent Unit (GRU) approach. The suggested bidirectional GRU layer is a computationally effective framework used in lightweight DCNNs to enhance overall system efficiency which also has a significant role in speech-to-text recognition. This work employed the standard CHiME-5 corpus comprises multi-speaker conversational recordings across many real-home dinner parties to assess the proposed framework. The word error rate (WER) of 34.65% was obtained by applying this dataset to the proposed bidirectional GRU-based DCNN framework. Also the suggested model performance is validated using the TED-LIUM andLibriSpeech datasets. Our proposed model achieved 10 .65% WER on the TED-LIUM dataset and 10.08% WER on LibriSpeech, respectively, which is a relative improvement in performance over baseline models.