Elamur Research⌕
☰
home / papers / 2026 / MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages

MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages

Paper · 2026 · Qiongqiong Wang, Ai Ti Aw, Nancy F. Chen et al.

Summary

The paper tackles binary speaker gender recognition (female/male) from speech, focusing on English and Southeast Asian languages, where existing tools generalize unevenly. The authors build MERaLiON-GR on the MERaLiON-SpeechEncoder-2 conformer backbone. They adapt it cheaply with LoRA in the attention projections, mix all 25 hidden layers through learned attention weights, and add an ECAPA-TDNN head with attention pooling and a small linear classifier. Training combines gender-labelled VoxCeleb1 with five IMDA partitions of Singapore's National Speech Corpus, which include read, conversational, code-switched and stylised speech. The model is evaluated on 15 public test sets in eight languages and on a new human-checked Singapore in-the-wild set, SG-ECMT. It beats Vox-Profile on 12 of the 15 public sets and on every in-the-wild set, and a general Audio-LLM lags far behind both. Supplying the predicted gender to the Audio-LLM as metadata greatly raises its accuracy. The authors conclude that dedicated paralinguistic models remain necessary and can also serve as signal providers for general multimodal systems.

Key points

  • Architecture: MERaLiON-SpeechEncoder-2 with 24 layers and LoRA (rank and scale 16, rsLoRA, dropout 0.05), soft attention over layer outputs, then an ECAPA-TDNN (512-dim, three SE-Res2Net blocks, GroupNorm replacing BatchNorm), a 256-dim embedding, and an RMSNorm+GELU linear head.
  • Training setup:
    • Each batch of 32 comes from a single dataset, chosen with probability proportional to dataset size.
    • The loss is class-balanced weighted cross-entropy (effective-number weights) with label smoothing 0.1.
    • The LoRA adapters and the head get separate learning rates.
    • Training ran for 15 epochs on one H100.
    • The final model averages the four best dev checkpoints.
  • Public benchmarks:
    • MERaLiON-GR reaches 100% on English FLEURS, Tamil OpenSLR and Thai Elderly.
    • Its largest gains over Vox-Profile are on Tamil EmoTa (+4.71 pp), Vietnamese Common Voice (+3.14 pp) and IndoWave (+3.00 pp).
    • It ties on Khmer FLEURS.
    • It loses on SMALDUSC (−4.40 pp) and THAI-SER (−2.09 pp), which the authors attribute to domain mismatch.
  • SG-ECMT construction:
    • The set contains about 470 samples each in Singlish, Chinese, Malay and Tamil.
    • Labels were first produced by a WavLM-ECAPA model using sliding windows.
    • Native-speaker annotators then corrected them.
    • On short segments, MERaLiON-GR improves on Vox-Profile by up to 4.32 pp (Malay).
  • Metadata injection: giving the predicted gender to MERaLiON-2-10B lifts its accuracy sharply, for example by +60.00 pp on Vietnamese Common Voice and +47.70 pp on English FLEURS.

Interesting ideas

  • I1. Instead of using only the final encoder layer, the model learns a soft weighting over every layer's output, so that phonetic and speaker cues from different depths can both contribute.
  • I2. Each batch is taken from one corpus only, and corpora are drawn in proportion to their size, so differently styled sources stay separated within an update while overall data usage stays fair.
  • I3. Averaging the weights of the several best dev checkpoints gives some of the robustness of an ensemble at no extra inference cost.
  • I4. The evaluation set was built by letting an existing model propose labels and then having native speakers correct them, which is much cheaper than labelling from scratch.
  • I5. A general model answered far better when a specialist's prediction was handed to it as explicit metadata, rather than being expected to infer that attribute itself.

For Mova

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

  • I1 (tree): gparse could learn a per-task mix of transformer layers for its arc and label scorers, and the learned weights would also show which depths the parser relies on.
  • I2 (learn): when training on mixed treebanks plus silver data from witness parsers, single-source batches could reduce interference between UD dialects and annotation styles.
  • I3 (tree): the perceptron parser and gparse could average their top-k checkpoints instead of keeping only the single best one.
  • I4 (learn): silver parses agreed by several witnesses could be turned into a small human-corrected gold test set, giving evaluation that does not depend on the silver pipeline itself.
  • I5 (answer): readers and LLM teacher tiers could receive explicit facts from the world stage (entities, states, event order) as structured context, which should be both more accurate and easier to explain.

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

🔬Institute of High Performance ComputingSingapore · Singapore5 works · 169 citations

Also on the byline: Institute of Advanced Intelligence and Computing.

 Fields, topics and concepts

📚 Shelves 5The 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 worksSignal Processing · 452 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,263 works

Cite

Cite as Qiongqiong Wang, Ai Ti Aw, Nancy F. Chen, Ying Lay Chiu, Yang Ding, Yingxu He, Ridong Jiang, Zhuohan Liu, Yanfeng Lu, Yi Ma, Muhammad Huzaifah, Nabilah Binte Md Johan, Nattadaporn Lertcheva, Pham Minh Duc, Sailor Hardik Bhupendra, Siti Umairah Binte Mohammad Salleh, Shuo Sun, Tarun Kumar Vangani, Jeremy H. M. Wong, Jinyang Wu and Longyin Zhang. 2026. MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages. arXiv:2608.04433.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2608.04433 · arXiv 2608.04433

@misc{wang2026meralion,
  title = {{MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages}},
  author = {Qiongqiong Wang, Ai Ti Aw, Nancy F. Chen, Ying Lay Chiu, Yang Ding, Yingxu He, Ridong Jiang, Zhuohan Liu, Yanfeng Lu, Yi Ma, Muhammad Huzaifah, Nabilah Binte Md Johan, Nattadaporn Lertcheva, Pham Minh Duc, Sailor Hardik Bhupendra, Siti Umairah Binte Mohammad Salleh, Shuo Sun, Tarun Kumar Vangani, Jeremy H. M. Wong, Jinyang Wu and Longyin Zhang},
  year = {2026},
  doi = {10.48550/arxiv.2608.04433},
  eprint = {2608.04433},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2608.04433}
}

References

27 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), pages 4218–4222, 2020. ↗
  2. A. Baevski, H. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems (NeurIPS), 2020.
  3. C. Busso, M. Bulut, C.-C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan. IEMOCAP: Interactive emotional dyadic motion capture database. Language Resources and Evaluation, 42(4):335–359, 2008. ↗
  4. A. Bustamin, A. M. Rizky, E. Warni, I. S. Areni, and I. Indrabayu. IndoWaveSentiment: Indonesian audio dataset for emotion classification. Data in Brief, 57:111138, 2024. ↗
  5. S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, et al. WavLM: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6):1505–1518, 2022. find
  6. C.-C. Chiu, J. Qian, Y. Zhang, A. Han, C. Zhang, E. Battenberg, and A. Natsev. Self-supervised learning with random-projection quantizer for speech and audio representation. In Proceedings of the International Conference on Machine Learning (ICML), 2022. find
  7. J. S. Chung, A. Nagrani, and A. Zisserman. VoxCeleb2: Deep speaker recognition. In Proc. Interspeech, pages 1086–1090, 2018. find
  8. A. Conneau, M. Ma, S. Khanuja, Y. Zhang, V. Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna. FLEURS: Few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 798–805.
  9. Y. Cui, M. Jia, T.-Y. Lin, Y. Song, and S. Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9268–9277, 2019. find
  10. Data Wow and VISAI. Thai elderly speech dataset. https://github.com/VISAI-DATAWOW/ Thai-Elderly-Speech-dataset, 2023. ↗
  11. B. Desplanques, J. Thienpondt, and K. Demuynck. ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification. In Proc. Interspeech, pages 3830–3834, 2020. find
  12. T. Feng, J. Lee, A. Xu, Y. Lee, T. Lertpetchpun, X. Shi, H. Wang, T. Thebaud, L. Moro-Velazquez, D. Byrd, et al. Vox-profile: A speech foundation model benchmark for characterizing diverse speaker and speech traits. arXiv preprint arXiv:2505.14648, 2025. ↗
  13. F. He, S.-H. C. Chu, O. Kjartansson, C. Rivera, A. Katanova, A. Gutkin, I. Demirsahin, C. Johny, M. Jansche, S. Sarin, and K. Pipatsrisawat. Open-source multi-speaker speech corpora for building Gujarati, Kannada, Malayalam, Marathi, Tamil and Telugu speech synthesis systems. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), pages 6494–6503, Marseille, France, 2020. European Language Resources Association (ELRA). find
  14. Y. He, Z. Liu, S. Sun, B. Wang, W. Zhang, X. Zou, N. F. Chen, and A. T. Aw. MERaLiON-AudioLLM: Technical report. arXiv preprint arXiv:2412.09818, 2024. ↗
  15. W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed. HuBERT: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451–3460, 2021. ↗
  16. E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022.
  17. D. Kalajdzievski. A rank stabilization scaling factor for fine-tuning with LoRA. arXiv preprint arXiv:2312.03732, 2023. ↗
  18. J. X. Koh, A. Mislan, K. Khoo, B. Ang, W. Ang, C. Ng, and Y.-Y. Tan. Building the Singapore English national speech corpus. In Proc. Interspeech, pages 321–325, 2019. doi: 10.21437/ Interspeech.2019-1525. find
  19. MagicHub. ASR-SMalDuSC: A scripted Malay daily-use speech corpus. https://magichub.com/ datasets/malay-scripted-speech-corpus-daily-use-sentence/, 2023. ↗
  20. MERaLiON Team. Meralion-speechencoder: Towards a speech foundation model for singapore and beyond, 2024. URL https://arxiv.org/abs/2412.11538. ↗
  21. A. Nagrani, J. S. Chung, and A. Zisserman. VoxCeleb: A large-scale speaker identification dataset. In Proc. Interspeech, pages 2616–2620, 2017. doi: 10.21437/Interspeech.2017-950. ↗
  22. J. Thevakumar, L. Thavarasa, T. Sivatheepan, S. Kugarajah, and U. Thayasivam. EmoTa: A Tamil emotional speech dataset. In Proceedings of the First Workshop on Challenges in Processing South Asian Languages (CHiPSAL 2025), Abu Dhabi, UAE, 2025. International Committee on Computational Linguistics. find
  23. B. Wang, X. Zou, S. Sun, W. Zhang, Y. He, Z. Liu, C. Wei, N. F. Chen, and A. Aw. Advancing Singlish understanding: Bridging the gap with datasets and multimodal models. arXiv preprint arXiv:2501.01034, 2025a. ↗
  24. Q. Wang, H. B. Sailor, T. Liu, and A. T. Aw. Contextual paralinguistic data creation for multi-modal speech-llm: Data condensation and spoken QA generation. In Proc. Interspeech, 2025b. find
  25. Q. Wang, H. B. Sailor, T. Liu, W. Zhang, M. Huzaifah, N. Lertcheva, S. Sun, N. F. Chen, J. Wu, and A. Aw. Benchmarking contextual and paralinguistic reasoning in speech-llms: A case study with in-the-wild data. In Findings of EMNLP 2025, 2025c. ↗
  26. Q. Wang, H. B. Sailor, J. H. M. Wong, T. Liu, S. Sun, W. Zhang, M. Huzaifah, N. Chen, and A. T. Aw. Incorporating contextual paralinguistic understanding in large speech-language models. In Proc. IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2025d. find
  27. J. Wongpithayadisai, C. Chaksangchaichot, S. Sangnark, P. Prakrankamanant, K. Gangwanpongpun, S. Boonpunmongkol, P. Milindasuta, D. Na-Pombejra, S. Nutanong, and E. Chuangsuwanich. THAI speech emotion recognition (THAI-SER) corpus. arXiv preprint arXiv:2507.09618, 2025. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-meralion-gr-speech-gender-recognition/