Speech Recognition System based on Wavelet Multi- Resolution Analysis using One-Dimensional CNN-LSTM Network

Speech Emotion Recognition (SER) is responsible for identifying the speaker’s emotions through speech and has a significant role in psychological assessment and human-computer interaction (HCI). Various time representations like Mel Spectrograms, spectrograms, as well as Mel Frequency Cepstral Coefficients (MFCCs) are widely utilized to develop SER systems. This representation uses Fast Fourier Transform (FFT) to translate the time domain signal into a frequency. However, FFT is constrained by the uncertainty principle and cannot attain great resolution in both time and frequency at the same time. In contrast, wavelets are able to offer better localization in both cases with their high resolution. Use autoencoders and combine long short term memory (LSTM) networks with one-dimensional convolutional neural networks (CNN). The scenes from the one- dimensional CNN-LSTM model are classified employing the latent space that results from the reduction of the autoencoder of the wavelet features' dimensionality. The method achieved an unweighted accuracy (UA) of 81.45% and a weighted accuracy (WA) of 81.22% when applied to the Ryerson Audiovisual Emotional Speech and Song Database (RAVDESS) dataset utilizing Monte-Carlo K-fold validation. The state- of-the-art method uses another time-frequency representation of the situation. Speech signal recognition is emerging research in human-computer interaction, and its uses incorporate human-computer interaction, the usability of virtual reality, behavioral assessment, and medical and emergency services. A novel method introduces an AI- assisted deep Stochastic convolutional neural network (DSCNN) architecture, which utilizes convolutional networks to enhance and learn important features in speech spectrograms. This model subsamples the feature maps with specific steps in the convolutional layers, bypassing the need for pooling layers and learning general features across all layers. Then, the SoftMax classifier is used for classification. Evaluation of Interactive Emotional Dyadic Motion Capture (IEMOCAP) and Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) data shows 7.85% and 4.5% accuracy and 34.5 MB reduced standard respectively, which shows the effectiveness and practicality of SER technology. Applications appear in SER due to its useful information. However, incorrect extraction rules and unclear solutions may limit the performance. To solve these problems, the psychoacoustic model inspired by speech coding introduces the information of the segmentation line to obtain a more comprehensive solution. Three new spectral characteristics— spectral flatness, spectral slope, and spectral entropy—are put forth. The hypothesis set is identified utilizing a support vector machine (SVM) classifier. Experiments indicate that this approach is more effective than other cutting-edge Fourier and multi-resolution amplitude features as well as MFCC features.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.