A Reproducible Hybrid LSTM–SVM Framework for Speech Emotion Recognition Using Temporal–Acoustic Embeddings

Abstract

Speech emotion recognition (SER) seeks to infer affective states from paralinguistic cues embedded in voice, but reliable evaluation is difficult because emotional speech is sequential, speaker-dependent, and sensitive to corpus and validation design. This study presents a hybrid Long Short-Term Memory–Support Vector Machine (LSTM–SVM) framework in which 40 MFCCs with first- and second-order deltas are combined with pitch, RMS energy, zero-crossing rate, spectral centroid, bandwidth, roll-off, and spectral-contrast descriptors to form frame-level acoustic sequences. The LSTM learns temporal embeddings and an RBF-kernel SVM performs final maximum-margin classification. The reconstructed experimental protocol uses the speech-only RAVDESS corpus: 1,440 utterances from 24 professional actors (12 female, 12 male) spanning eight emotions (neutral, calm, happy, sad, angry, fearful, disgust, and surprised). Evaluation is specified as five-fold speaker-independent grouped cross-validation, with normalization and augmentation fitted or applied only within the training partition. The surviving aggregate evaluation record reports 84.60% accuracy, 84.20% macro-precision, 83.80% macro-recall, and 84.00% macro-F1 for the LSTM–SVM model, compared with 76.80% accuracy for standalone LSTM and 70.40% for standalone SVM. Removing MFCCs reduced accuracy to 74.20%, while removing pitch and energy reduced it to 79.50%. Because the original workstation storage failed after the experiment, fold-level logs and per-utterance predictions could not be recovered; therefore, no unrecoverable significance statistic or confusion matrix is presented as measured evidence. The revised contribution is a transparent, leakage-aware architecture and a fully specified reconstruction protocol that separates temporal representation learning from margin-based classification.