Scholarlink

<p>Arabic speech command recognition in industrial environments remains challenging due to dialectal diversity, limited training data, and high acoustic variability. Feature representation and model design play a central role in the performance of automatic speech recognition (ASR) systems. In this paper, we investigate three time–frequency representations: MFCCs, classical spectrograms, and mel-spectrograms. Experiments show that mel-spectrograms provide the best baseline accuracy (92.50%) compared to MFCCs (91.00%) and spectrograms (84.33%). To improve performance, we apply a multi-level data augmentation strategy including pitch shifting, time stretching, and SpecAugment, as well as a temporal attention mechanism within a CNN–LSTM framework. Data augmentation gives the largest gain, reaching 97.17% accuracy. Temporal attention improves temporal modeling and achieves 95.67%. The combination of augmentation and attention yields 96.50%, offering a balanced trade-off between accuracy and stability. The Arabic Industrial Command Corpus, developed for this work, contains 3,000 samples from 30 speakers across 10 commands with balanced phonetic coverage. These findings demonstrate the value of combining perceptually motivated features, data augmentation, and attention to enhance the robustness of Arabic ASR in noisy industrial conditions.</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

Arabic speech command recognition using an enhanced CNN-LSTM model with attention and data augmentation

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Naouar Laaidi,
  • Meryam Telmem,
  • Mohamed Lamrini,
  • Hassan Satori

摘要

Arabic speech command recognition in industrial environments remains challenging due to dialectal diversity, limited training data, and high acoustic variability. Feature representation and model design play a central role in the performance of automatic speech recognition (ASR) systems. In this paper, we investigate three time–frequency representations: MFCCs, classical spectrograms, and mel-spectrograms. Experiments show that mel-spectrograms provide the best baseline accuracy (92.50%) compared to MFCCs (91.00%) and spectrograms (84.33%). To improve performance, we apply a multi-level data augmentation strategy including pitch shifting, time stretching, and SpecAugment, as well as a temporal attention mechanism within a CNN–LSTM framework. Data augmentation gives the largest gain, reaching 97.17% accuracy. Temporal attention improves temporal modeling and achieves 95.67%. The combination of augmentation and attention yields 96.50%, offering a balanced trade-off between accuracy and stability. The Arabic Industrial Command Corpus, developed for this work, contains 3,000 samples from 30 speakers across 10 commands with balanced phonetic coverage. These findings demonstrate the value of combining perceptually motivated features, data augmentation, and attention to enhance the robustness of Arabic ASR in noisy industrial conditions.