Paper deep dive
MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages
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, Longyin Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 8/9/2026, 3:46:54 AM
Summary
The paper introduces MERaLiON-GR, a speech gender recognition system for English and Southeast Asian languages. It fine-tunes the MERaLiON-SpeechEncoder-2 backbone using Low-Rank Adaptation (LoRA) and appends an ECAPA-TDNN downstream network. The model outperforms state-of-the-art baselines like Vox-Profile and MERaLiON-v2 (Audio-LLM) across multiple benchmarks and languages, demonstrating robust cross-lingual generalization and improved performance when gender metadata is injected into Audio-LLMs.
Entities (34)
Relation Signals (29)
MERaLiON-GR → performstask → Gender Recognition
confidence 95% · speech gender recognition system that performs binary classification (female / male)
MERaLiON-GR → supportslanguage → Indonesian
confidence 95% · Extensive evaluations across multilingual Singaporean and Southeast Asian languages (...Indonesian...
MERaLiON-GR → supportslanguage → Thai
confidence 95% · Extensive evaluations across multilingual Singaporean and Southeast Asian languages (...Thai...
MERaLiON-GR → supportslanguage → Vietnamese
confidence 95% · Extensive evaluations across multilingual Singaporean and Southeast Asian languages (...Vietnamese...
MERaLiON-GR → supportslanguage → Khmer
confidence 95% · Extensive evaluations across multilingual Singaporean and Southeast Asian languages (...and Khmer)
MERaLiON-GR → supportslanguage → English
confidence 95% · performs binary classification (female / male) on English and Southeast Asian (SEA) languages.
MERaLiON-GR → supportslanguage → Chinese
confidence 95% · Extensive evaluations across multilingual Singaporean and Southeast Asian languages (English, Chinese...
MERaLiON-GR → supportslanguage → Malay
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present MERaLiON-GR, a speech gender recognition system that performs binary classification (female / male) on English and Southeast Asian (SEA) languages. The model finetunes MERaLiON-SpeechEncoder-2, a large conformer based transformer pre-trained on a broad speech corpus, and applies parameter efficient fine-tuning via Low-Rank Adaptation (LoRA) to adapt the encoder to the gender recognition task, and appends a multi-scale ECAPA-TDNN down stream network with attention pooling and a lightweight linear classifier. Extensive evaluations across multilingual Singaporean and Southeast Asian languages (English, Chinese, Malay, Tamil, Thai, Vietnamese, Indonesian, and Khmer) show that MERaLiON-GR consistently surpasses the state-of-the-art gender recognition model Vox-Profile and a large Audio-LLM, in both full-utterance and segment level evaluation modes. The results underscore the value of dedicated speech models in achieving accurate paralinguistic understanding and strong cross-lingual generalization.
Tags
Links
- Source: https://arxiv.org/abs/2608.04433v1
- Canonical: https://arxiv.org/abs/2608.04433v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
25,356 characters extracted from source content.
Expand or collapse full text
MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages MERaLiON Team Institute of Advanced Intelligence and Computing (IAIC), A*STAR, Singapore Corresponding Authors:Qiongqiong Wang (wang_qiongqiong@a-star.edu.sg) Abstract We present MERaLiON-GR111The MERaLiON-GR-v1 model: https://huggingface.co/MERaLiON/MERaLiON-GR-v1. An online demo: https://meralion.org/analysis. , a speech gender recognition system that performs binary classification (female / male) on English and Southeast Asian (SEA) languages. The model finetunes MERaLiON-SpeechEncoder-2, a large conformer-based transformer pre-trained on a broad speech corpus, and applies parameter-efficient fine-tuning via Low-Rank Adaptation (LoRA) to adapt the encoder to the gender recognition task, and appends a multi-scale ECAPA-TDNN downstream network with attention pooling and a lightweight linear classifier. Extensive evaluations across multilingual Singaporean and Southeast Asian languages (English, Chinese, Malay, Tamil, Thai, Vietnamese, Indonesian, and Khmer) show that MERaLiON-GR consistently surpasses the state-of-the-art gender recognition model Vox-Profile and a large Audio-LLM, in both full-utterance and segment-level evaluation modes. The results underscore the value of dedicated speech models in achieving accurate paralinguistic understanding and strong cross-lingual generalization. 1 Introduction Gender recognition from speech is a fundamental paralinguistic task with applications ranging from speaker diarization and speaker profiling to personalised speech interfaces and audio forensics. While human listeners can reliably infer speaker gender from brief utterances, automatic systems must generalize across diverse recording conditions, languages, accents, age groups, and channel characteristics. In contrast to earlier approaches based on handcrafted acoustic features, such as fundamental frequency (F0), formants, and spectral descriptors, modern deep learning methods learn discriminative representations directly from speech signals. These methods typically employ convolutional, recurrent, ECAPA-TDNN (Desplanques et al., 2020), or Transformer-based architectures operating on raw waveforms or log-Mel filterbank features. More recently, self-supervised speech models, including HuBERT (Hsu et al., 2021), wav2vec 2.0 (Baevski et al., 2020), and WavLM (Chen et al., 2022), have significantly advanced speech representation learning. Their pretrained representations encode rich acoustic, speaker, and linguistic information, enabling substantial improvements across downstream paralinguistic tasks, including gender recognition. In this work, we built a complete pipeline for multilingual speech gender recognition based on a large pre-trained conformer encoder, MERaLiON-SpeechEncoder-2222The MERaLiON-SpeechEncoder-2 model: https://huggingface.co/MERaLiON/MERaLiON-SpeechEncoder-2 (MERaLiON Team, 2024). MERaLiON-SpeechEncoder-2 is a large conformer-based transformer trained with a Best-RQ objective (Chiu et al., 2022) on a multilingual corpus that includes a significant proportion of Singapore English and other Southeast Asian languages, making it particularly well suited for gender recognition benchmarks centred on Southeast Asia. However, fully fine-tuning all 24 layers is computationally expensive and may increase the risk of catastrophic forgetting. To address this, we adopt Low-Rank Adaptation (LoRA) (Hu et al., 2022), which parameterizes weight updates in the attention projection layers using low-rank factors rather than updating the full weight matrices. This yields a parameter-efficient adaptation strategy for the downstream task. The main contributions of this work are: (1) a parameter-efficient multilingual GR model built on MERaLiON-SpeechEncoder-2 with LoRA and ECAPA-TDNN, outperforming Vox-Profile (Feng et al., 2025) on 12 of 15 public benchmarks spanning eight languages across Southeast Asia and English; (2) robust segment-level recognition on short in-the-wild speech, with gains of up to 4.32 p over Vox-Profile on Singapore language sets; and (3) a demonstration that injecting GR estimates as paralinguistic metadata substantially improves Audio-LLM performance on gender-related question answering, consistent with our prior findings on contextual paralinguistic understanding in Audio-LLMs (Wang et al., 2025c). 2 Model description The overall architecture consists of three components: (1) the pre-trained MERaLiON-SpeechEncoder-2 backbone (MERaLiON Team, 2024), (2) a layer-attentive aggregation module, and (3) an ECAPA-TDNN downstream network followed by a linear classification head. 2.1 Backbone: MERaLiON-SpeechEncoder-2 with LoRA MERaLiON-SpeechEncoder-2 is a Conformer-based transformer (MERaLiON Team, 2024). To enable parameter-efficient adaptation while reducing computational cost and the risk of overfitting, we insert Low-Rank Adaptation (LoRA) (Hu et al., 2022) adapters into the attention projection layers. The LoRA rank and the scaling factor are both set to 16. We further adopt rsLoRA normalization (Kalajdzievski, 2023) and apply a LoRA dropout rate of 0.05. All 25 hidden-state outputs (the input embeddings plus all 24 transformer layers) are retained. A learned soft-attention mechanism over per-layer mean representations produces a weighted combination H=∑lalhlH= _la_lh_l, capturing complementary phonetic and speaker-level cues from different layers. 2.2 Downstream Model: ECAPA-TDNN The aggregated representation is processed by an ECAPA-TDNN network adapted for gender recognition (Desplanques et al., 2020). It is projected to 512 dimensions and passed through three SE-Res2Net blocks (dilations 1,2,3\1,2,3\, SE reduction factor 8); All BatchNorm layers are replaced with GroupNorm layers that are in the original ECAPA-TDNN structure. Attention pooling aggregates temporal frames into a fixed-length embedding e, which is projected to 256 dimensions and classified by a linear head with RMSNorm and GELU activation. The classification head maps the 256-dimensional embedding to two class logits: y^=W2g(RMSNorm(W1)) y=W_2\,g\! (RMSNorm(W_1e) ) (1) where g(⋅)g(·) denotes the GELU activation. 3 Training Details 3.1 Training datasets The training data consists of six dataset partitions: VoxCeleb1 (Nagrani et al., 2017) (gender-labelled) and IMDA PART1–5 (Singaporean English). • VoxCeleb1: Segments drawn from the VoxCeleb1 corpus. • IMDA PART1–5: We use five partitions of a large-scale Singapore speech corpus developed by the Info-communications Media Development Authority (IMDA) of Singapore and derived from the National Speech Corpus (NSC) (Koh et al., 2019; Wang et al., 2025a). They cover a broad range of speaking conditions: prompted phonetically-balanced scripts with local accents (PART1); prompted sentences drawn from everyday topics such as food, people, and locations, rich in local terms (PART2); spontaneous conversational speech on daily-life topics (PART3); code-switching between Singapore English and Mandarin, Malay, or Tamil (PART4); and stylised speech including debates, finance discussions, and emotional expressions (PART5). This diversity of speaking styles, accents, and code-switching patterns makes the NSC particularly valuable for training a robust multilingual Gender recognition model targeting Southeast Asian speech. 3.2 Training Configuration A multi-dataset training strategy was adopted across the six training partitions described in Section 3.1. To ensure balanced optimization across datasets of different sizes, each training batch consisted of 32 samples drawn from a single dataset, where the dataset was sampled with probability proportional to its size. Class imbalance was addressed using the class-balanced weighted cross-entropy loss described in Section 3.2. To improve model robustness without incurring the inference overhead of model ensembling, the final model parameters were obtained by averaging the weights of the four checkpoints with the lowest categorical loss on the development set. Optimization employed two parameter groups: the LoRA adapters inserted into the frozen backbone were trained with a learning rate of (5×10−55× 10^-5) and a weight decay of (5×10−45× 10^-4), while the downstream classification network used a learning rate of (6×10−46× 10^-4) and a weight decay of (8×10−58× 10^-5). A cosine learning-rate scheduler with a linear warm-up ratio of 0.08 was applied throughout training. Label smoothing with (ϵ=0.1)ε=0.1) was used to improve generalization. The model was trained for 15 epochs on a single NVIDIA H100 GPU, using the development-set categorical loss as the early stopping criterion. To address class imbalance, we optimize the model using a weighted cross-entropy loss for gender classification: ℒCE=−∑i=1Cwi⋅yilog(yi^),L_CE=- _i=1^Cw_i· y_i ( y_i), (2) where C denotes the number of gender classes, yiy_i and yi y_i are the ground truth and the predicted probability for the i-th class, respectively. wiw_i is the weight assigned to the i-th class to compensate for class imbalance. The per-class weights wiw_i are computed from the effective number of training samples per class (Cui et al., 2019). 4 Evaluation Setup 4.1 Evaluation Datasets Model performance was assessed using both manually curated Singapore speech datasets and publicly available multilingual benchmarks, providing evaluation across a wide range of languages and recording conditions. 4.1.1 Manually Curated Singapore Evaluation Set The Singapore evaluation set (SG-ECMT) comprises speech samples in the country’s four major languages: Singlish, Mandarin Chinese, Malay, and Tamil. The data was derived from our proprietary unlabeled speech corpus and consists of 10–30 second in-the-wild speech segments. Following the data processing pipeline proposed in (Wang et al., 2025b), speech segments were selected from a large collection of in-the-wild recordings based on their paralinguistic metadata (Wang et al., 2025d). Gender labels were automatically estimated using a WavLM-ECAPA333https://github.com/wenet-e2e/wespeaker model (Chen et al., 2022; Desplanques et al., 2020) fine-tuned on the VoxCeleb2 dataset (Chung et al., 2018). Inference was performed using a 3-second sliding window with a 1-second overlap (i.e., a 2-second hop), producing one gender prediction every two seconds. The predicted gender labels were subsequently reviewed and corrected by native-speaking annotators for each language to ensure annotation quality. The final evaluation set comprises 466 Singlish, 466 Chinese, 479 Malay, and 469 Tamil speech samples. 4.1.2 Public evaluation set To evaluate generalization, we leverage a diverse collection of public benchmarks spanning eight languages: English (FLEURS (Conneau et al., 2023), IEMOCAP (Busso et al., 2008), and Common Voice (Ardila et al., 2020)), Chinese (Common Voice (Ardila et al., 2020)), Malay (SMALDUSC (MagicHub, 2023)), Tamil (OpenSLR SLR65 (He et al., 2020), EmoTa (Thevakumar et al., 2025), and Common Voice (Ardila et al., 2020)), Thai (THAI-SER (Wongpithayadisai et al., 2025), Thai Elderly Speech (Data Wow and VISAI, 2023), and Common Voice (Ardila et al., 2020)), Vietnamese (Common Voice (Ardila et al., 2020)), Indonesian (IndoWaveSentiment (Bustamin et al., 2024) and Common Voice (Ardila et al., 2020)), and Khmer (FLEURS (Conneau et al., 2023)). 5 Baselines and Metrics We benchmark MERaLiON-GR against two representative baselines. Vox-Profile (Feng et al., 2025) is a speech foundation model designed to characterize diverse speaker and speech attributes and represents the current state of the art in standalone gender recognition. MERaLiON-v2444The MERaLiON-2-10B model: https://huggingface.co/MERaLiON/MERaLiON-2-10B is a general-purpose Audio-LLM built on the MERaLiON-AudioLLM framework (He et al., 2024), capable of performing natural-language instruction following over speech input. We include it as a representative Audio-LLM baseline to evaluate whether a general-purpose multimodal model can perform fine-grained paralinguistic recognition without task-specific training. For the standalone gender recognition models (Vox-Profile and the proposed MERaLiON-GR), performance is evaluated using classification accuracy. To ensure a fair comparison, MERaLiON-v2 is prompted with a fixed instruction: PROMPT = "Determine the speaker’s gender in the given audio. Reply with a single label from: Female, Male" The generated response is mapped to one of the two gender labels, and classification accuracy is computed using the same evaluation protocol as the standalone gender recognition models. 6 Results and Discussion 6.1 Public Benchmark Results Table 1 reports gender recognition (GR) accuracy on the public multilingual benchmarks. We compare the proposed MERaLiON-GR model with the state-of-the-art standalone GR model, Vox-Profile, and the general-purpose Audio-LLM, MERaLiON-v2. Overall, MERaLiON-v2 consistently underperforms the dedicated GR models, indicating that a general-purpose Audio-LLM is less effective for fine-grained gender recognition without task-specific optimization. Compared with Vox-Profile, MERaLiON-GR achieves higher accuracy on 12 of the 15 evaluated test sets and matches its performance on the Khmer FLEURS benchmark. The proposed model attains perfect accuracy (100.00%) on the English FLEURS, Tamil OpenSLR, and Thai Elderly benchmarks. The largest improvements are observed on Tamil EmoTa (+4.71,p), Vietnamese Common Voice (+3.14,p), and Indonesian IndoWave (+3.00,p), demonstrating strong generalization across diverse languages and recording conditions. Performance decreases are observed only on the Malay SMALDUSC dataset ( −4.40-4.40,p) and the Thai SER dataset (−2.09-2.09,p). These degradations are likely due to domain mismatch, as both datasets differ substantially from the training data in terms of recording conditions and speaking style. Table 2 investigates whether incorporating the predicted gender from MERaLiON-GR can improve Audio-LLM performance. Specifically, the predicted gender is injected as paralinguistic metadata together with the input audio (GR-AudioLLM). Compared with the Audio-LLM, GR-AudioLLM consistently improves accuracy across all evaluated datasets, with particularly large gains on Vietnamese Common Voice (+60.00,p), English FLEURS (+47.70,p), Tamil Common Voice (+37.72,p), and Malay SMALDUSC (+41.20,p). These results demonstrate that explicitly providing gender information substantially enhances the Audio-LLM’s ability to answer gender-related paralinguistic questions. This observation is consistent with our previous findings (Wang et al., 2025b), where injecting paralinguistic metadata similarly improved contextual paralinguistic question answering (CPQA), suggesting that explicit paralinguistic cues provide an effective conditioning signal for speech-language understanding. Table 1: Accuracy (%) on public datasets. Lang. Test Set Vox-Profile MERaLiON-GR Audio-LLM English FLEURS 99.69 100.00 49.61 IEMOCAP 97.31 98.90 97.21 Common Voice 92.60 93.90 52.10 Chinese Common Voice 96.10 98.10 64.80 Malay SMALDUSC 97.60 93.20 57.10 Tamil OpenSLR 98.30 100.00 53.70 EmoTa 94.44 99.15 69.98 Common Voice 92.30 94.00 51.60 Thai THAI-SER 89.32 87.23 79.05 Thai Elderly 96.57 100.00 72.18 Common Voice 96.52 97.86 59.97 Vietnamese Common Voice 96.08 99.22 36.08 Indonesian IndoWave 95.33 98.33 71.00 Common Voice 95.00 97.40 50.02 Khmer FLEURS 99.74 99.74 69.80 Table 2: Accuracy (%) on public datasets with paralinguistic metadata. Lang. Test Set Audio-LLM MERaLiON-GR GR-AudioLLM English FLEURS 49.61 100.00 97.31 IEMOCAP 97.21 98.90 97.97 Malay SMALDUSC 57.10 93.20 98.30 Tamil EmoTa 69.98 99.15 92.30 Common Voice 51.60 94.00 89.32 Thai Common Voice 59.97 97.86 72.90 Vietnamese Common Voice 36.08 99.22 96.08 Khmer FLEURS 69.80 99.74 99.74 6.2 In-the-Wild Results To evaluate robustness in unconstrained environments, we conduct 2-second segment-level evaluations on our internal in-the-wild recordings, consisting of 10–30-second utterances in four Singapore languages (Singlish, Chinese, Malay, and Tamil). Table 3 shows that the proposed model consistently outperforms Vox-Profile across all evaluation sets. The improvements are particularly evident on these short, in-the-wild utterances, with gains of up to 4.32,p. The largest improvement is achieved on Malay (+4.32,p), where Vox-Profile also exhibits its lowest baseline accuracy, followed closely by Tamil (+4.16,p). These results demonstrate that our model provides robust gender predictions for fine-grained, segment-level pseudo-labeling and paralinguistic metadata generation, making it a practical replacement for external GR models. Table 3: Accuracy (%) on in-the-wild datasets (10–30 s segments). Lang. Test Set Vox-Profile MERaLiON-GR Chinese SG-ECMT-Chinese 89.25 91.85 Singlish SG-ECMT-Singlish 88.45 92.10 Malay SG-ECMT-Malay 86.56 90.88 Tamil SG-ECMT-Tamil 90.57 94.73 7 Summary We presented MERaLiON-GR, a speech gender recognition model for English and Southeast Asian languages. Extensive experiments on both manually curated Singapore datasets and public multilingual benchmarks demonstrate that MERaLiON-GR consistently outperforms existing baselines across diverse languages, recording conditions, and speech durations. Furthermore, incorporating the predicted gender as paralinguistic metadata significantly improves the performance of a general-purpose Audio-LLM on gender-related speech understanding tasks. These results demonstrate the continued importance of dedicated speech encoders with task-specific paralinguistic modeling for robust gender recognition and cross-lingual generalization. Acknowledgments and Disclosure of Funding This research/project is supported by the National Research Foundation, Singapore under its National Large Language Models Funding Initiative. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation, Singapore. The computational work for this article was fully performed on resources of the National Supercomputing Centre (NSCC), Singapore (https://w.nscc.sg). The authors would also like to thank Nattadaporn Lertcheva, Nabilah Binte Md Johan, Amudha Narayanan, Anitha Veeramani, Siti Maryam Binte Ahmad Subaidi, Siti Umairah Md Salleh, James Tan, and Komathee Veerappan to perform human annotations for the gender annotation task. References R. Ardila, M. Branson, K. Davis, M. Kohler, J. Meyer, M. Henretty, R. Morais, L. Saunders, F. Tyers, and G. Weber (2020) Common voice: a massively-multilingual speech corpus. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC), p. 4218–4222. Cited by: §4.1.2. A. Baevski, H. Zhou, A. Mohamed, and M. Auli (2020) wav2vec 2.0: a framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems (NeurIPS), Cited by: §1. C. Busso, M. Bulut, C. Lee, A. Kazemzadeh, E. Mower, S. Kim, J. N. Chang, S. Lee, and S. S. Narayanan (2008) IEMOCAP: interactive emotional dyadic motion capture database. Language Resources and Evaluation 42 (4), p. 335–359. Cited by: §4.1.2. A. Bustamin, A. M. Rizky, E. Warni, I. S. Areni, and I. Indrabayu (2024) IndoWaveSentiment: Indonesian audio dataset for emotion classification. Data in Brief 57, p. 111138. Cited by: §4.1.2. S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, et al. (2022) WavLM: large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing 16 (6), p. 1505–1518. Cited by: §1, §4.1.1. C. Chiu, J. Qian, Y. Zhang, A. Han, C. Zhang, E. Battenberg, and A. Natsev (2022) Self-supervised learning with random-projection quantizer for speech and audio representation. In Proceedings of the International Conference on Machine Learning (ICML), Cited by: §1. J. S. Chung, A. Nagrani, and A. Zisserman (2018) VoxCeleb2: deep speaker recognition. In Proc. Interspeech, p. 1086–1090. Cited by: §4.1.1. A. Conneau, M. Ma, S. Khanuja, Y. Zhang, V. Axelrod, S. Dalmia, J. Riesa, C. Rivera, and A. Bapna (2023) FLEURS: few-shot learning evaluation of universal representations of speech. In 2022 IEEE Spoken Language Technology Workshop (SLT), p. 798–805. Cited by: §4.1.2. Y. Cui, M. Jia, T. Lin, Y. Song, and S. Belongie (2019) Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9268–9277. Cited by: §3.2. Data Wow and VISAI (2023) Thai elderly speech dataset. Note: https://github.com/VISAI-DATAWOW/Thai-Elderly-Speech-dataset Cited by: §4.1.2. B. Desplanques, J. Thienpondt, and K. Demuynck (2020) ECAPA-TDNN: emphasized channel attention, propagation and aggregation in tdnn based speaker verification. In Proc. Interspeech, p. 3830–3834. Cited by: §1, §2.2, §4.1.1. T. Feng, J. Lee, A. Xu, Y. Lee, T. Lertpetchpun, X. Shi, H. Wang, T. Thebaud, L. Moro-Velazquez, D. Byrd, et al. (2025) Vox-profile: a speech foundation model benchmark for characterizing diverse speaker and speech traits. arXiv preprint arXiv:2505.14648. Cited by: §1, §5. F. He, S. C. Chu, O. Kjartansson, C. Rivera, A. Katanova, A. Gutkin, I. Demirsahin, C. Johny, M. Jansche, S. Sarin, and K. Pipatsrisawat (2020) 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), Marseille, France, p. 6494–6503. Cited by: §4.1.2. Y. He, Z. Liu, S. Sun, B. Wang, W. Zhang, X. Zou, N. F. Chen, and A. T. Aw (2024) MERaLiON-AudioLLM: technical report. arXiv preprint arXiv:2412.09818. Cited by: §5. W. Hsu, B. Bolte, Y. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed (2021) HuBERT: self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM Transactions on Audio, Speech, and Language Processing 29, p. 3451–3460. Cited by: §1. E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen (2022) LoRA: low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), Cited by: §1, §2.1. D. Kalajdzievski (2023) A rank stabilization scaling factor for fine-tuning with LoRA. arXiv preprint arXiv:2312.03732. Cited by: §2.1. J. X. Koh, A. Mislan, K. Khoo, B. Ang, W. Ang, C. Ng, and Y. Tan (2019) Building the Singapore English national speech corpus. In Proc. Interspeech, p. 321–325. External Links: Document, ISSN 2958-1796 Cited by: 2nd item. MagicHub (2023) ASR-SMalDuSC: a scripted Malay daily-use speech corpus. Note: https://magichub.com/datasets/malay-scripted-speech-corpus-daily-use-sentence/ Cited by: §4.1.2. MERaLiON Team (2024) MERaLiON-speechencoder: towards a speech foundation model for singapore and beyond. External Links: 2412.11538, Link Cited by: §1, §2.1, §2. A. Nagrani, J. S. Chung, and A. Zisserman (2017) VoxCeleb: a large-scale speaker identification dataset. In Proc. Interspeech, p. 2616–2620. External Links: Document Cited by: §3.1. J. Thevakumar, L. Thavarasa, T. Sivatheepan, S. Kugarajah, and U. Thayasivam (2025) EmoTa: a Tamil emotional speech dataset. In Proceedings of the First Workshop on Challenges in Processing South Asian Languages (CHiPSAL 2025), Abu Dhabi, UAE. Cited by: §4.1.2. B. Wang, X. Zou, S. Sun, W. Zhang, Y. He, Z. Liu, C. Wei, N. F. Chen, and A. Aw (2025a) Advancing Singlish understanding: bridging the gap with datasets and multimodal models. arXiv preprint arXiv:2501.01034. Cited by: 2nd item. Q. Wang, H. B. Sailor, T. Liu, and A. T. Aw (2025b) Contextual paralinguistic data creation for multi-modal speech-llm: data condensation and spoken QA generation. In Proc. Interspeech, Cited by: §4.1.1, §6.1. Q. Wang, H. B. Sailor, J. H. M. Wong, T. Liu, S. Sun, W. Zhang, M. Huzaifah, N. Chen, and A. T. Aw (2025c) Incorporating contextual paralinguistic understanding in large speech-language models. In Proc. IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), Cited by: §1. Q. Wang, H. B. Sailor, T. Liu, W. Zhang, M. Huzaifah, N. Lertcheva, S. Sun, N. F. Chen, J. Wu, and A. Aw (2025d) Benchmarking contextual and paralinguistic reasoning in speech-llms: a case study with in-the-wild data. In Findings of EMNLP 2025, Cited by: §4.1.1. J. Wongpithayadisai, C. Chaksangchaichot, S. Sangnark, P. Prakrankamanant, K. Gangwanpongpun, S. Boonpunmongkol, P. Milindasuta, D. Na-Pombejra, S. Nutanong, and E. Chuangsuwanich (2025) THAI speech emotion recognition (THAI-SER) corpus. arXiv preprint arXiv:2507.09618. Cited by: §4.1.2. 8 MERaLiON Team (alphabetical order) Aw Ai Ti, Chen Fang Yih Nancy, Chiu Ying Lay, Ding Yang, He Yingxu, Jiang Ridong, Liu Zhuohan, Lu Yanfeng, Ma Yi, Muhammad Huzaifah Bin Md Shahrin, Nabilah Binte Md Johan, Nattadaporn Lertcheva, Pham Minh Duc, Sailor Hardik Bhupendra, Siti Umairah Binte Mohammad Salleh, Sun Shuo, Tarun Kumar Vangani, Wang Qiongqiong, Wong Heng Meng Jeremy, Wu Jinyang, Zhang Longyin