IMPLEMENTATION OF ROBUST SPEECH RECOGNITION TECHNIQUES FOR IMPAIRED SPEECH PATTERNS

Abstract

These Recent advances in speech recognition have made possible precise transcription and command recognition in a variety of application areas. Nevertheless, the current systems have a major problem in identifying speech of afflicted people, including the patients with dysarthria, post-stroke, or growth speech disorders. To overcome this deficiency, the current paper presents a new framework that will help enhance the performance of speech recognition among impaired speakers. The system first divides the speech into male and female to be able to use the specifics of the voice and the differences in acoustics. After that, a large data set of impaired and healthy samples of speech is prepared and processed. Noise reduction and acoustic data extraction, such as Mel-Frequency Cepstral Coefficients (MFCCs) and Mel-spectrograms, are part of the preprocessing phase. The core design concept in the effective use of CNNs and LSTM networks is their combination to achieve spatial and temporal speech patterns recording. Once the first transcription has been made, the identified text (impaired) is fixed with the help of an n-gram language model to strengthen its contextual interpretation and linguistic coherence. To verify the strength of the proposed model, the Word Error Rate (WER) is used to evaluate it and compare it with deep learning architectures. As per the experimental results, the proposed methodology is much better than the existing speech recognition methods in terms of accuracy and ability to withstand damaged speech. Therefore, the system improves human-computer interaction among affected people and leads to the creation of more inclusive and adaptive speech recognition systems.