This study on isolated word recognition in the Malayalam language examines the application of the Wavelet Scattering Transform (WST) in conjunction with Convolutional Neural Networks (CNN). While WST have demonstrated success in speech recognition across multiple languages, their application to Malayalam speech recognition remains unexplored. The proposed hybrid approach is expected to enhance the recognition accuracy by leveraging the strengths of both the WST and CNN methodologies by tackling the unique challenges associated with isolated word recognition in the context of Malayalam language. Also, wavelet scattering transform is highly beneficial for small datasets. The utilization of WST with CNN leads to a 12% enhancement in recognition accuracy compared to Mel - Frequency Cepstral Coefficients (MFCC) with CNN, and an 11% improvement with Mel spectrogram using CNN. The F1 score also reflected a similar improvement. Experimental results demonstrate the effectiveness of the proposed system in achieving accurate and efficient recognition of Malayalam isolated words.

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

Isolated Word Recognition in Malayalam Using the Wavelet Scattering Transform and CNN

  • Fathima Kunhi Mohamed,
  • P. P. Abdul Haleem

摘要

This study on isolated word recognition in the Malayalam language examines the application of the Wavelet Scattering Transform (WST) in conjunction with Convolutional Neural Networks (CNN). While WST have demonstrated success in speech recognition across multiple languages, their application to Malayalam speech recognition remains unexplored. The proposed hybrid approach is expected to enhance the recognition accuracy by leveraging the strengths of both the WST and CNN methodologies by tackling the unique challenges associated with isolated word recognition in the context of Malayalam language. Also, wavelet scattering transform is highly beneficial for small datasets. The utilization of WST with CNN leads to a 12% enhancement in recognition accuracy compared to Mel - Frequency Cepstral Coefficients (MFCC) with CNN, and an 11% improvement with Mel spectrogram using CNN. The F1 score also reflected a similar improvement. Experimental results demonstrate the effectiveness of the proposed system in achieving accurate and efficient recognition of Malayalam isolated words.