Scholarlink

<p>This study proposes an end-to-end Automatic Speech Recognition (ASR) system for Punjabi based on a Convolutional Neural Network (CNN) architecture that directly processes raw speech signals to improve recognition accuracy. Unlike conventional ASR methods, the model integrates tonal features—including pitch, loudness, and intensity—via a fully connected layer to better capture the tonal nuances of the language. The CNN layers extract acoustic features from raw speech, and their outputs are fused with tonal information before classification through a Softmax layer. The system was trained and tested on a comprehensive Punjabi speech corpus containing 28&#xa0;h of recordings and 64,600 utterances annotated with tonal characteristics. Experimental results reveal that the model achieves its best performance with two convolutional layers and that incorporating tonal features lowers the Word Error Rate to 10.7%, significantly outperforming existing state-of-the-art Punjabi ASR systems. These findings highlight the critical role of tonal information in enhancing speech recognition for tonal languages like Punjabi.</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

Tonality-Aware CNN Model for Punjabi Automatic Speech Recognition

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Ahmed Alkhayyat,
  • Halijah Hassan,
  • K. N. Raja Praveen,
  • Mandeep Kaur Chohan,
  • Shivakrishna Dasi,
  • Ankur Srivastava,
  • Sankara Rao Palla,
  • Devendra Singh

摘要

This study proposes an end-to-end Automatic Speech Recognition (ASR) system for Punjabi based on a Convolutional Neural Network (CNN) architecture that directly processes raw speech signals to improve recognition accuracy. Unlike conventional ASR methods, the model integrates tonal features—including pitch, loudness, and intensity—via a fully connected layer to better capture the tonal nuances of the language. The CNN layers extract acoustic features from raw speech, and their outputs are fused with tonal information before classification through a Softmax layer. The system was trained and tested on a comprehensive Punjabi speech corpus containing 28 h of recordings and 64,600 utterances annotated with tonal characteristics. Experimental results reveal that the model achieves its best performance with two convolutional layers and that incorporating tonal features lowers the Word Error Rate to 10.7%, significantly outperforming existing state-of-the-art Punjabi ASR systems. These findings highlight the critical role of tonal information in enhancing speech recognition for tonal languages like Punjabi.