Machine Learning-Based Speech and Noise Classification in Cocktail Party Environments Using NMF and PSO-Optimized Feature Selection
Year: 2026
Type: conference-paper
Abstract: The problem of cocktail party is a serious problem in the processing of speech where speech signals have to be identified and categorized amid conflicting noise signals. This paper suggests a full-fledged machine learning system to identify both speech and noise in noisy acoustics. The system commences with preprocessing clean speech and noise signals and then mixing the summation of the signals with different signal-to-noise ratio (SNR) to recreate realistic situations. We then apply Non-negative Matrix Factorization (NMF) to decompose the mixed signals to approximate speech and noise signals. Acoustic features including Mel-Frequency Cepstral Coefficients (MFCCs) and spectral descriptors are obtained out of the separated signals. Particle Swarm Optimization (PSO) is used to select features to improve the performance of the models and lower redundancy of features. Various machine learning models are trained with the selected features, such as Support Vector Machine with RBF kernel (SVM-RBF), Random Forest, and XGBoost. Experimental findings reveal that the SVM-RBF model is the most successful with an accuracy of 86.25% and F1-score of 0.862 in comparison to XGBoost and Random Forest. The findings reveal that signal separation, feature optimization, and machine learning are effective in speech classification in cocktail party settings.
Institutions: Northern Technical University, Istanbul Aydın University
Language: English
Topic: Speech and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7225182425