Scholarlink

<p>This study proposes a deep learning (DL)-based approach for Kashmiri isolated speech recognition using mel-spectrogram features and a hybrid architecture integrating one-dimensional convolutional neural networks (CNNs) with gated recurrent units (GRUs). The model was evaluated on a newly curated dataset comprising 3,630 spoken Kashmiri numeral samples across ten distinct classes. Preprocessing techniques, such as silence trimming and noise reduction, were applied, followed by mel-spectrogram extraction for feature representation. The hybrid architecture is formulated based on insights from a comprehensive ablation study aimed at optimizing model design for Kashmiri speech recognition (SR). The CNN component extracts spatial features, while the GRU module models temporal dependencies. A thorough ablation study was conducted to evaluate the influence of different hyperparameters on model performance. The hybrid model achieved a performance with 90.77% accuracy using a configuration of three Conv1D layers (64 kernels, 5<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\({\times }\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>5 filter size), three average-pooling layers, two GRU layers (128 units each), and two dense layers (32 units each) employing PReLU activation. A dropout rate of 0.3 was applied uniformly across CNN, GRU, and dense layers. Training employed a batch size of 16, with Kullback–Leibler (KL) divergence as the loss function and the Adam optimizer configured with a learning rate of 0.001. Results underscore the effectiveness of the hybrid CNN-GRU architecture in capturing both spatial and temporal patterns in mel-spectrogram features, surpassing standalone CNN and GRU models, and demonstrating promise for SR in low-resource languages like Kashmiri. Furthermore, a comparative analysis with conventional machine learning techniques revealed the higher performance of the proposed approach.</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

Developing Hybrid One-dimensional CNN-GRU Model for Kashmiri Speech Recognition

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Muzaffar Ahmad Dar,
  • Jagalingam Pushparaj

摘要

This study proposes a deep learning (DL)-based approach for Kashmiri isolated speech recognition using mel-spectrogram features and a hybrid architecture integrating one-dimensional convolutional neural networks (CNNs) with gated recurrent units (GRUs). The model was evaluated on a newly curated dataset comprising 3,630 spoken Kashmiri numeral samples across ten distinct classes. Preprocessing techniques, such as silence trimming and noise reduction, were applied, followed by mel-spectrogram extraction for feature representation. The hybrid architecture is formulated based on insights from a comprehensive ablation study aimed at optimizing model design for Kashmiri speech recognition (SR). The CNN component extracts spatial features, while the GRU module models temporal dependencies. A thorough ablation study was conducted to evaluate the influence of different hyperparameters on model performance. The hybrid model achieved a performance with 90.77% accuracy using a configuration of three Conv1D layers (64 kernels, 5 \({\times }\) × 5 filter size), three average-pooling layers, two GRU layers (128 units each), and two dense layers (32 units each) employing PReLU activation. A dropout rate of 0.3 was applied uniformly across CNN, GRU, and dense layers. Training employed a batch size of 16, with Kullback–Leibler (KL) divergence as the loss function and the Adam optimizer configured with a learning rate of 0.001. Results underscore the effectiveness of the hybrid CNN-GRU architecture in capturing both spatial and temporal patterns in mel-spectrogram features, surpassing standalone CNN and GRU models, and demonstrating promise for SR in low-resource languages like Kashmiri. Furthermore, a comparative analysis with conventional machine learning techniques revealed the higher performance of the proposed approach.