EmoNet speech-emotion CNN reports 96.75% accuracy with 4.11M parameters
Original title: EmoNet: A Lightweight Attention-Based CNN for Speech Emotion Recognition
The model classifies seven emotions on combined TESS, RAVDESS and SAVEE data using a separable CNN and BiLSTM.
Researchers propose EmoNet, a lightweight speech emotion recognition model for seven emotion categories. It uses a four-block depthwise separable CNN with squeeze-and-excitation attention and residual connections, a bidirectional LSTM, and multi-head attention pooling, plus a 134-channel acoustic feature set and augmentation including SpecAugment. On a combined TESS, RAVDESS and SAVEE set, the authors report 96.75% test accuracy with 4.11 million parameters, beating their reproduced baseline and staying competitive with larger models.
Unconfirmed: The 96.75% accuracy is the authors' reported test result on their combined dataset split.
- Training uses audio-level perturbations and SpecAugment.
- The encoder has four depthwise separable CNN blocks with residual connections.
- Authors say it remains competitive with much larger transformer-based systems.
- Evaluation covers the combined TESS, RAVDESS and SAVEE datasets.
From the original article
1 min read at arxiv.orgSpeech Emotion Recognition (SER) is a fundamental task in affective computing that enables applications such as mental health assessment, human-computer interaction, and automated customer support systems. Accurate emotion classification from speech remains challenging due to variations in speaker style, acoustic conditions, and the high computational cost of modern deep learning approaches.
By Muhammad Moiz Khalid · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.