Scholarlink

<p>In Northern India, underscoring the need for Automatic Speech Recognition (ASR) systems specifically tailored to regional language. Most existing ASR systems are primarily developed for English and other European languages, typically utilizing feature extraction techniques such as Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP), which may not perform effectively in real-world scenarios. This study investigates and compares various feature extraction methods to determine the most suitable approach for Punjabi speech recognition under both clean and noisy conditions. It also evaluates the performance of different acoustic models, with particular emphasis on Context-Dependent (CD) models. Experimental results demonstrate that MFCC features achieve a Word Error Rate (WER) of 12% in clean environments, while Gammatone Frequency Cepstral Coefficients (GFCCs) result in a WER of 14.8% in noisy environments, both using a Bidirectional Long Short-Term Memory (BLSTM) model.</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

Feature Extraction and Context-Dependent Acoustic Modeling for Robust Speech Recognition

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Ahmed Alkhayyat,
  • S. Narayanasamy,
  • Ibrahim Oteir,
  • Deeksha Verma,
  • K. N. Raja Praveen,
  • Sarbeswara Hota,
  • Rajesh Singh,
  • Saurabh Namdev

摘要

In Northern India, underscoring the need for Automatic Speech Recognition (ASR) systems specifically tailored to regional language. Most existing ASR systems are primarily developed for English and other European languages, typically utilizing feature extraction techniques such as Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction (PLP), which may not perform effectively in real-world scenarios. This study investigates and compares various feature extraction methods to determine the most suitable approach for Punjabi speech recognition under both clean and noisy conditions. It also evaluates the performance of different acoustic models, with particular emphasis on Context-Dependent (CD) models. Experimental results demonstrate that MFCC features achieve a Word Error Rate (WER) of 12% in clean environments, while Gammatone Frequency Cepstral Coefficients (GFCCs) result in a WER of 14.8% in noisy environments, both using a Bidirectional Long Short-Term Memory (BLSTM) model.