Scholarlink

<p>In primary school children, delay in speech and language development has been observed recently more than before. It is essential to detect the delay early and provide speech therapy to improve the quality of speech. Doctors can provide recommendations to avoid speech delay and enhance the child’s development. The objective of the work is to detect speech impairment in children to enhance their development. Another objective is to develop a single speech impairment detection module that can detect impairment in all type of speech such as vowels, consonants, one/two/three syllables, difficult words, etc. The inputs used to detect speech impairment is speech recordings from children. The pitch features are then extracted from speech signals called as CHROMA features. The pitch features are then given as input to recurrent neural network called stacked LSTM network to detect speech impairment. The results of the speech impairment detection method is analyzed in terms of precision, recall, AUC, F1-score and accuracy. The F1-score is 1 and accuracy is 99.99% for speech impairment detection for different type of speech. The F1-score is 0.999, and the accuracy of the method is found to be 99.99% for combined speech impairment detection. The speech impairment detection has been also compared with MFCC feature based method and it is found that CHROMA features are detecting speech impairment more accurately. Advantage of the method is that it can detect all type of impaired speech using a single module.</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

A novel method to identify speech impairment in children using pitch features and stacked LSTM recurrent neural network

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • Manisa Manoswini,
  • Aleena Swetapadma,
  • Biswajit Sahoo

摘要

In primary school children, delay in speech and language development has been observed recently more than before. It is essential to detect the delay early and provide speech therapy to improve the quality of speech. Doctors can provide recommendations to avoid speech delay and enhance the child’s development. The objective of the work is to detect speech impairment in children to enhance their development. Another objective is to develop a single speech impairment detection module that can detect impairment in all type of speech such as vowels, consonants, one/two/three syllables, difficult words, etc. The inputs used to detect speech impairment is speech recordings from children. The pitch features are then extracted from speech signals called as CHROMA features. The pitch features are then given as input to recurrent neural network called stacked LSTM network to detect speech impairment. The results of the speech impairment detection method is analyzed in terms of precision, recall, AUC, F1-score and accuracy. The F1-score is 1 and accuracy is 99.99% for speech impairment detection for different type of speech. The F1-score is 0.999, and the accuracy of the method is found to be 99.99% for combined speech impairment detection. The speech impairment detection has been also compared with MFCC feature based method and it is found that CHROMA features are detecting speech impairment more accurately. Advantage of the method is that it can detect all type of impaired speech using a single module.