End-to-End recognition approach for Cognitive Impaired speech using Sequential Conv-Nets
- Resource Type
- Conference
- Authors
- Sharma, Riya; Gupta, Saloni; Gambhir, Pooja; Bansal, Poonam
- Source
- 2022 4th International Conference on Artificial Intelligence and Speech Technology (AIST) Artificial Intelligence and Speech Technology (AIST), 2022 4th International Conference on. :1-6 Dec, 2022
- Subject
- Computing and Processing
Robotics and Control Systems
Training
Time-frequency analysis
Neural networks
Speech enhancement
Software
Convolutional neural networks
Spectrogram
Automatic speech recognition (ASR)
Convolutional Neural Network (CNN)
Sequential End-to-End approach
Mel-Filters
Mild Cognitive Impairment Disorder
- Language
It is essential to find a way to early diagnosis of Mental Cognitive Disorders (MCD) in order to enable preventative care and prompt therapy to stop future progression. Automatic speech recognition software (ASR) giving transcriptions could possibly enhance communication of speech in the present. ASR is especially beneficial for those with increasing conditions that limit the comprehensibility of speech issues with motor function. ASR services typically have training on normal speech and might not be ideal for speech impairment, putting up a roadblock to using augmented help tools. This paper presents the recognition of cognitive impaired speech using Sequential 2-dimensional Conv-Nets. Convolutional networks efficiently explore and exploits temporal, cepstral and spectral structures of the speech signals. The network has been trained and tested on Spectrograms of Mel-Filters banks taking its time and frequency variants. The experiment was performed on 15000 spoken responses of digits (0-9) uttered by Dementia patients. The outcome of the trained model showed a validation accuracy of 91.27% with a loss of 0.5%.