Comparative Evaluation of Random Forest and 1D CNN Using MFCC Features for Embedded Voice-Command Recognition

Abstract

Comparative studies of Random Forest (RF) and Convolutional Neural Network (CNN) classifiers for voice-command recognition typically report only which model is more accurate, without explaining why the weaker model underperforms or assessing whether either model can be deployed on a resource-constrained microcontroller. This study addresses that gap by comparing RF and a one-dimensional CNN for Mel-Frequency Cepstral Coefficient (MFCC)-based voice classification, diagnosing the specific cause of RF's underperformance, and evaluating estimated deployment feasibility on an ESP32-S3 microcontroller. Using 15,400 augmented recordings from the Google Speech Commands dataset across eleven command classes, both models were evaluated on a dedicated test set. An RF classifier evaluated on time-averaged MFCC features reached only 49% test accuracy; the confusion matrix, feature-importance, principal component analysis, and centroid-distance analyses converged on one explanation: averaging across time erases the temporal structure of speech, collapsing acoustically similar classes into overlapping feature-space regions. Retaining the complete, un-averaged MFCC sequence raised the RF test accuracy to 73%, but the resulting model far exceeded a typical microcontroller's flash budget. Conversely, a CNN trained on the identical full-sequence representation reached 93% test accuracy. Based on hardware profiling estimates, this CNN requires only a kilobyte-scale flash memory footprint, replacing the massive multi-megabyte arrays required by RF. These results demonstrate that preserving temporal structure is critical for both classification accuracy and compact model size in embedded voice-command recognition. Future work will address current study limitations by incorporating silence and unknown background classes, and by ensuring strictly speaker-independent evaluation protocols for robust real-world deployment.