Elamur Research⌕
☰
home / papers / 2025 / ArabEmoNet: A Lightweight Hybrid 2D CNN-BiLSTM Model with Attention for Robust Arabic Speech Emotion Recognition

ArabEmoNet: A Lightweight Hybrid 2D CNN-BiLSTM Model with Attention for Robust Arabic Speech Emotion Recognition

Paper · 2025 · Ali Abouzeid, Bilal Elbouardi, Mohamed Maged et al.

Summary

Arabic speech emotion recognition (SER) has little labelled data. Earlier Arabic systems mostly fed 13 MFCC coefficients into 1D convolutions, which handle time and frequency structure separately. The authors present ArabEmoNet, a compact model with about 0.97M parameters. It turns raw audio into Log-Mel spectrograms and runs one pipeline over them: 2D convolutions, then a two-layer BiLSTM, then attention pooling over time, then a classifier. During training, Gaussian noise and SpecAugment-style augmentation help against overfitting.

The model was tested on two Arabic emotion corpora, KSUEmotions and KEDAS. On KSUEmotions it reaches 91.48% accuracy, about 4 points above the previous best CNN-BLSTM-DNN system. On KEDAS it reaches 99.46%. Both results match or beat fine-tuned HuBERT-base and Whisper models that are dozens of times larger.

Ablations support three design choices:

  • 7×7 kernels work best.
  • Augmentation adds a little over 2 points.
  • Replacing the BiLSTM with a Transformer lowers accuracy.

The work matters because it shows that a small, well-designed model can compete with large pretrained ones when data is scarce and compute is limited.

Key points

  • Input features: Log-Mel spectrograms with 128 Mel bands over 80–7600 Hz, an FFT window of 2048, a hop of 256 and a Hann window, converted to dB relative to the maximum power.
  • Architecture:
    • three 2D conv stages with 32/64/128 filters, 7×7 kernels, 2×2 max pooling and dropout 0.3;
    • a 2-layer BiLSTM with 64 units per direction;
    • additive attention over the 128-dim BiLSTM outputs.
  • Results on KSUEmotions: 91.48% accuracy and 91.46% macro F1.
    • Strongest baseline, CNN-BLSTM-DNN (Hifny and Ali): 87.20%.
    • HuBERT-base (95M parameters): 87.04%.
    • The CNN-Transformer variant: 86.66%.
  • Results on KEDAS: 99.46% accuracy, against a 75.00% original baseline and 99.40% for Whisper-small.
  • Ablations on KSUEmotions:
    • Kernel size: 3 gives 89.71%, 7 peaks at 91.48% with 0.97M parameters, and 11 drops to 89.90% with 1.71M parameters.
    • Augmentation raises accuracy from 89.10% to 91.48%.
    • Per-emotion accuracy is lowest for Happy (88.37%) and highest for Fear (96.92%).
  • Stated limitation: shorter clips are zero-padded to the longest clip in the batch. This wastes computation and may affect learning when clip lengths vary a lot.

Interesting ideas

  • I1. Inspectable attention pooling: Instead of averaging the sequence, the model learns one scalar weight per time step and builds a weighted summary, so one can see which frames drove the decision.
  • I2. Report model size next to accuracy: Every comparison lists parameter count beside the score, which makes "small but nearly as good" a measurable claim rather than a slogan.
  • I3. Swap one component, freeze the rest: The BiLSTM is replaced with a Transformer while everything else stays fixed, which isolates what the temporal encoder itself contributes.
  • I4. Sweep one hyperparameter and track cost: The kernel-size sweep reveals a clear sweet spot: too small misses context, too large adds noise and parameters.
  • I5. Cheap noise augmentation with an on/off ablation: Adding random noise during training is simple, and the paper proves its value by reporting results with and without it.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (frames): Clause- or token-level weights for irony or speech-act cues could be shown as part of the explanation of why a sentence was read as ironic.
  • I2 (learn): Evaluation tables comparing the perceptron parser with gparse could always show parameters and latency next to LAS, so the cost of each gain is visible.
  • I3 (tree): When testing gparse encoder variants, freeze tokenisation, features and decoding so differences in LAS can be attributed to the encoder alone.
  • I4 (tokens): A similar sweep over the tagger's context-window width, recording both accuracy and feature count, could find the cheapest window that still disambiguates tags.
  • I5 (learn): Inject controlled noise such as typos, casing changes and dropped punctuation into silver training data, and always report the matching no-noise run as a control.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of WaterlooWaterloo · Canada67 works · 13,592 citations🎓Mohamed bin Zayed University of Artificial IntelligenceAbu Dhabi · United Arab Emirates124 works · 5,676 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 worksSignal Processing · 452 works

🧭 Topics 3Research subjects and languages the work is about.Sentiment & emotion · 972 worksSemitic languages · 2,908 worksSpeech & spoken language · 480 works

Cite

Cite as Ali Abouzeid, Bilal Elbouardi, Mohamed Maged and Shady Shehata. 2025. ArabEmoNet: A Lightweight Hybrid 2D CNN-BiLSTM Model with Attention for Robust Arabic Speech Emotion Recognition. Proceedings of The Third Arabic Natural Language Processing Conference.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.arabicnlp-main.17

@inproceedings{abouzeid2025arabemonet,
  title = {{ArabEmoNet: A Lightweight Hybrid 2D CNN-BiLSTM Model with Attention for Robust Arabic Speech Emotion Recognition}},
  author = {Ali Abouzeid, Bilal Elbouardi, Mohamed Maged and Shady Shehata},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference},
  year = {2025},
  doi = {10.18653/v1/2025.arabicnlp-main.17},
  url = {https://doi.org/10.18653/v1/2025.arabicnlp-main.17}
}

References

23 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR. ↗
  2. Mourad Belhadj, Ilham Bendellali, and Elalia Lakhdari. 2022. Kedas: A validated arabic speech emotion dataset. In 2022 International Symposium on iNnovative Informatics of Biskra (ISNIB), pages 1–6. IEEE. find
  3. Hazem M Fayek, Marcin Lech, and Luis Cavedon. 2017. Evaluating deep learning architectures for speech emotion recognition. Neural Networks, 92:60–68. ↗
  4. Yasser Hifny and Ahmed Ali. 2019a. Efficient arabic emotion recognition using deep neural networks. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6710–6714. ↗
  5. Youssef Hifny and Ahmed Ali. 2019b. Efficient arabic emotion recognition using deep neural networks. In ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 6710–6714. IEEE. ↗
  6. Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. 2021. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. In IEEE/ACM Transactions on Audio, Speech, and Language Processing, volume 29, pages 3451–3460. IEEE. ↗
  7. Mina Huh, Ruchira Ray, and Corey Karnei. 2024. A comparison of speech data augmentation methods using s3prl toolkit. Preprint, arXiv:2303.00510. ↗
  8. Rashid Jahangir, Ying Wah Teh, Faiqa Hanif, and Ghulam Mujtaba. 2021. Deep learning approaches for speech emotion recognition: state of the art and research challenges. Multimedia Tools and Applications, 80(16):23745–23812. ↗
  9. Waleed Akram Khan, Hamad ul Qudous, and Asma Ahmad Farhan. 2024. Speech emotion recognition using feature fusion: a hybrid approach to deep learning. Multimedia Tools and Applications, 83(31):75557– 75584. find
  10. Nattapong Kurpukdee, Tomoki Koriyama, Takao Kobayashi, Sawit Kasuriya, Chai Wutiwiwatchai, and Poonlap Lamsrichan. 2017. Speech emotion recognition using convolutional long short-term memory neural network and support vector machines. In 2017 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pages 1744–1749. find
  11. E. Lieskovska, M. Jakubec, R. Jarina, and M. Chmulik. 2021. A review on speech emotion recognition using deep learning and attention mechanism. Electronics, 10:1163. ↗
  12. Ali Hamid Meftah, Mustafa A. Qamhan, Yasser Seddiq, Yousef A. Alotaibi, and Sid Ahmed Selouani. 2021. King saud university emotions corpus: Construction, analysis, evaluation, and comparison. IEEE Access, 9:54201–54219. ↗
  13. Seyed-Ahmad Mirsamadi, Eslam Barsoum, and Chao Zhang. 2017. Automatic speech emotion recognition using recurrent neural networks with local attention. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). find
  14. Swami Mishra, Nehal Bhatnagar, Prakasam P, and Sureshkumar T. R. 2024. Speech emotion recognition and classification using hybrid deep cnn and bilstm model. Multimedia Tools and Applications, 83(13):37603–37620. ↗
  15. Tin Lay Nwe, Say Wei Foo, and Liyanage C. De Silva. 2003. Speech emotion recognition using hidden markov models. Speech Communication, 41(4):603– 623. ↗
  16. David S Park and 1 others. 2019. Specaugment: A simple data augmentation method for automatic speech recognition. In INTERSPEECH, pages 2613–2617.
  17. SS Poorna, Vivek Menon, and Sundararaman Gopalan. 2025. Hybrid cnn-bilstm architecture with multiple attention mechanisms to enhance speech emotion recognition. Biomedical Signal Processing and Control, 100:106967. ↗
  18. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning (ICML). PMLR. ↗
  19. Tara N Sainath, Carolina Parada, Brian Kingsbury, and Bhuvana Ramabhadran. 2015. Convolutional, long short-term memory, fully connected deep neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). find
  20. Björn Schuller, Anton Batliner, Stefan Steidl, and Dino Seppi. 2011. Recognising realistic emotions and affect in speech: State of the art and lessons learnt from the first challenge. Speech communication, 53(9-10):1062–1087. ↗
  21. Marina Sokolova, Nathalie Japkowicz, and Stan Szpakowicz. 2009. A systematic analysis of performance measures for classification tasks. Information Processing & Management, 45(4):427–437. ↗
  22. George Trigeorgis, Marios A Nicolaou, Stefanos Zafeiriou, and Björn W Schuller. 2016. Adieu features? end-to-end speech emotion recognition using a deep convolutional recurrent network. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). ↗
  23. Ashish Vaswani and 1 others. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/abouzeid-2025-arabemonet-lightweight-hybrid-2d-cnn/