Komparasi CNN-BiLSTM, CRNN-BIGRU-Attention, dan Emotion2vec+ untuk Deteksi Emosi Suara Pelanggan

Abstract

Speech emotion recognition can support customer service by identifying users’ emotional states more quickly and consistently. This study compares CNN-BiLSTM, CRNN-BiGRU-Attention, and emotion2vec+ in classifying anger, sadness, happy, and neutral. The data comprise 5,764 selected recordings from RAVDESS and CREMA-D, processed separately using a speaker-independent split. CNN-BiLSTM and CRNN-BiGRU-Attention were trained from scratch using log-mel spectrograms, whereas emotion2vec+ performed direct inference on raw waveforms without fine-tuning. The evaluation covered clean audio, babble noise and white noise at SNR levels of 15, 10, and 5 dB, cross-corpus testing, false negative rate (FNR) for anger and sadness, and computational efficiency. emotion2vec+ achieved the highest clean-data Unweighted Average Recall (UAR), reaching 93.49% on RAVDESS and 91.52% on CREMA-D, average noisy-condition UARs of 82.68% and 71.67%, and maximum critical-emotion FNRs of 9.38% and 14.29%, respectively. The highest cross-corpus UAR was only 33.69%, indicating limited cross-corpus generalization. CRNN-BiGRU-Attention was the most efficient model, with 1,692,997 parameters, a 6.47 MB weight file, and an inference time of 1.3805-1.6569 milliseconds per sample. Thus, emotion2vec+ offered superior classification performance and noise robustness, whereas CRNN-BiGRU-Attention was preferable when computational efficiency was prioritized.