Enhancing Hindi Speech Recognition with Deep CNN-LiGRU through Mel and Gammatone Frequency Cepstral Coefficient Fusion

In India, Hindi stands as the most prevalent language. An ideal communication system in a local language would enable regular individuals to engage with machines via speech interfaces for information retrieval and daily tasks. Despite notable advancements in recent decades, achieving natural and robust human-machine speech interaction in regional languages remains challenging, especially in environments characterized by significant noise and reverberation..In such challenging, noisy environments, traditional automatic speech recognition systems employing Mel Frequency Cepstral Coefficients (MFCC) or Perceptual Linear Prediction (PLP) features have shown subpar performance. Research indicates that Gammatone Frequency Cepstral Coefficients (GFCC) perform well in noisy environments. To address this challenge, a novel approach combined features (MFGF) extracted from MFCC and GFCC to harness the strengths of both while mitigating their respective weaknesses. To bolster the robustness of speech recognition, contemporary speech recognition systems commonly leverage acoustic models based on Recurrent Neural Networks (RNNs), which inherently capture extensive temporal contexts and long-term speech modulations. The study introduces a Deep Convolutional- Light Gated Neural Network (CNN-LiGRU) model. This paper presents a refined iteration of Gated Recurrent Units (GRUs) termed Light Gated Neural Network (LiGRU), coupled with Convolution Neural Network (CNN).CNN extracts spatial features, capturing relevant patterns from the input spectrogram, while LiGRU handles the sequential nature of speech, capturing long-term dependencies and enhancing the model’s proficiency in understanding and recognizing spoken language. CNN-LiGRU demonstrates a 30% improvement in recognition rates over CNN-Recurrent Neural Network (RNN). Additionally, LiGRU eliminates the reset gate, resulting in reduced per-epoch training time compared to conventional RNNs. A comparative analysis with state-of-the-art methods is conducted, and experimental results indicate that the fusion approach outperforms, achieving speech recognition accuracy with minimal loss.