Optimizing Audio Recognition for Assistive Robotics with Feature Optimization, Machine Learning and Data Augmentation

Audio recognition plays a crucial role in assistive robotics, enabling intelligent systems to interpret and respond to various sound inputs. This study investigates the effectiveness of multiple feature extraction techniques (MFCC, MPEG-7, LPC, LPCC, PLP, and RASTA-PLP) with feature dimensions ranging from 10 to 38 (step of 2) and 64. These features were evaluated using 12 classifiers, including both traditional machine learning models (SVM, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k$</tex>-NN, GNB, LDA, RF, DT, SGD, MLP) and deep learning models (CNN, ANN), utilizing 10 -fold cross-validation repeated 10 times. To enhance performance, a dataset of 1,920 Romanian-language audio signals recorded in a simulated assistive environment was expanded to 4,440 samples by including both male and female voices. Further improvements were made using data augmentation techniques (noise injection, loudness variation, time and frequency masking, pitch shifting, and speed variation) resulting in a final dataset of 26,640 samples. The best accuracy was achieved using CNN with Mel-spectrograms at 97.86 %, though at a high computational cost (5,529.33 seconds). Meanwhile, SVM with MFCC-38 provided a competitive accuracy of 96.4 % with significantly lower processing time (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2,465.24$</tex> seconds). These findings highlight the trade-off between accuracy and computational efficiency, emphasizing CNN's superiority for precision and SVM's advantage in realtime applications. This study contributes to the development of robust audio recognition systems for assistive robotics, particularly for elderly care applications.