Speech Recognition using Wavelet Autoencoder and Convolutional Recurrent Neural Network
Abstract
Audio classification involves analyzing sound recordings and assigning them to specific categories. However, it faces several challenges, including unwanted noise, irregular phoneme articulation and scarcity or abundance of data due to reduced classification accuracy. This research proposed Wavelet Autoencoder and Convolutional Recurrent Neural Network (WA-CRNN) technique, which effectively addresses challenges and improves accuracy in audio classification. The WA-CRNN technique captures both spatial and temporal dependencies, allowing it to effectively learn from audio data. Initially, data collected from Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) and the pre-processing stage utilized noise removal, shifting and normalization to eliminate unwanted noise and scale the features, resulting in a uniform feature vector. The Mel-Frequency Cepstral Coefficient (MFCC) is used for feature extraction, which converts the audio signal to Mel scale, emphasizing the magnitude of frequency components. The wavelet-transformed input is passed through an autoencoder to reduce dimensionality, simplifying the input for CRNN while retaining essential features. The proposed model achieves an accuracy of 97.12% on the RAVDESS dataset and provides comparable performance to the existing deep learning approaches such as Convolutional Neural Networks (CNN), Stacked Hybrid Long Short-Term Memory (LSTM) technique and Transformer Encoder.