Automatic speech recognition systems are crucial tools that simplify communication. In particular, they are very helpful to hearing-impaired people. Such systems convert speech into written text, however they are still struggling with problems arising from noisy environments, accents, and complicated speech patterns. To address these issues, this work introduces a Denoise-Fused Adaptive Deep Network that is a hybrid 1D–2D convolutional architecture for correct speech digit recognition (0–9). Initially, the raw audio signals are obtained and the Savitzky–Golay filter (S–G filter) is used to de-noise them, which not only very efficiently removes high-frequency noise but also preserves the waveform. This pre-processing stage is a kind of insurance that the feature extraction stage will take place on clean and informative data thus recognition reliability will be elevated. The temporal and spectral aspects of the sounds are represented by Mel-Frequency Cepstral Coefficients (MFCC), Linear Prediction Cepstral Coefficients (LPCC), and spectrogram features taken from the clean signals. These features are fed to the Hybrid 1D–2D Adaptive Residual DenseNet (HC-ARDNet) that includes 1D convolutions for the sequential feature patterns and 2D convolutions for the spatial spectral structures. The presented network is made to operate at its peak by the use of Enhanced Egret Swarm Optimization (EESO) that changes the most important hyperparameters such as learning rate, kernel size, and the number of filters to their optimal values. The crux of the experiments is to provide evidence for the effectiveness of the proposed system, which has been accomplished by the figures of accuracy (96%), sensitivity (95%), precision (96%), and specificity (98%) that, basically, support the system as being not only stable but also capable of efficiently identifying spoken digits even when difficult are situations.
Tamilarasan et al. (Sun,) studied this question.