Speech Emotion Recognition via Knowledge Distillation Method Using CNN–BiLSTM-Transformer Pipeline

Read full text Summarise with VIRAT
Authors: Maather Alkhafaji, Amir Lakizadeh
Cited by: 0
References: 37

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

We propose a cross-lingual speech-emotion recognition framework that converts each utterance into a 128×128 log-Mel spectrogram, splits it into 64 32×8 patches, and feeds the tokens to a teacher with a Residual-CNN front end, a bidirectional LSTM, and a six-block Transformer encoder.A three-block student is trained by layer-wise feature-map distillation.Augmentations-additive noise and pitch shifts-regularize training across English and German corpora.Methodologically, we show that (i) patch tokenization with recurrent context yields languageagnostic suprasegmental representations without hand-crafted features; and (ii) Huber-based feature-map distillation transfers them from a six-layer teacher to a three-layer student, enabling an accuracy-efficiency trade-off for edge.On CREMA-D, EMO-DB, RAVDESS, SAVEE, and TESS, the teacher attains 94.07%, 96.50%, 96.89%, 98.85%, and 99.88% accuracy; the student achieves 93.56%, 95.96%, 96.53%, 98.53%, and 99.65% with 3.26M parameters and 0.41 GFLOPs-reductions of 41% and 75% over the 5.60M, 1.63-GFLOP teacher-supporting privacysensitive, low-power affect sensing.

Fields of Study & Subjects

Emotion and Mood Recognition Speech Recognition and Synthesis Speech and Audio Processing

Publication details

Journal / source
International journal of intelligent engineering and systems
Volume
19
Issue
4
Pages
830–846
Publication date
23 March 2026
Language
EN