MERaLiON/MERaLiON-GI-v1

MERaLiON-GR: Multi-lingual Speech Gender Recognition system that performs bi-nary classification (female / male) on English and Southeast Asian (SEA) languages.

• 🤗 Model: MERaLiON/MERaLiON-GI-v1
• 🎤 Live demo: meralion.org/analysis

---

🧠 Architecture Overview

• Modality: Speech-only model
• Task: Audio Classification (Gender Recognition)
• Input Sampling Rate: 16,000 Hz
• Output Classes: Female (0), Male (1)

Component	Description
Backbone	MERaLiON-SpeechEncoder-2 — 24-layer Conformer transformer, pre-trained with a Best-RQ objective
Adaptation	LoRA adapters in the attention projection layers (rank = 16, scaling factor = 16, rsLoRA normalization, dropout = 0.05)
Layer aggregation	Learned soft-attention over all 25 hidden-state outputs (input embeddings + 24 layer outputs)
Downstream network	ECAPA-TDNN — 3 SE-Res2Net blocks (dilations 1, 2, 3; SE reduction factor 8); BatchNorm replaced with GroupNorm
Pooling	Attention pooling → fixed-length speaker embedding, projected to 256-d
Classification head	Linear head with RMSNorm + GELU activation → 2-class output (Female / Male)

The design achieves parameter-efficient adaptation for multilingual, paralinguistic modeling.

More details on model architecture, training, and evaluation are available in the technical report: arxiv.org/abs/2608.04433

---

📊 Performance Overview

Public Benchmarks (Accuracy %)

Language	Test Set	Vox-Profile	MERaLiON-GR	Audio-LLM
English	FLEURS	99.69	100.00	49.61
English	IEMOCAP	97.31	98.90	97.21
English	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
Tamil	EmoTa	94.44	99.15	69.98
Tamil	Common Voice	92.30	94.00	51.60
Thai	THAI-SER	89.32	87.23	79.05
Thai	Thai Elderly	96.57	100.00	72.18
Thai	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
Indonesian	Common Voice	95.00	97.40	50.02
Khmer	FLEURS	99.74	99.74	69.80

MERaLiON-GR outperforms Vox-Profile on 12 of 15 public test sets and matches it on 1 set.

Singapore Language Performance

Language	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

In-the-wild 10–30 second Singapore speech segments (evaluated with a 2-second sliding window), MERaLiON-GR consistently outperforms Vox-Profile, with gains up to +4.32 pp (Malay) and +4.16 pp (Tamil).

Baselines

• Vox-Profile: https://huggingface.co/tiantiaf/wavlm-large-age-sex state-of-the-art standalone speech foundation model for gender recognition
• MERaLiON-v2 (MERaLiON-2-10B): general-purpose Audio-LLM, prompted with a fixed instruction to output a gender label

---

⚙️ Usage Examples

import torch
import torch.nn.functional as F
from transformers import AutoModel, AutoFeatureExtractor

# 1. Load the model and feature extractor
model_id = "MERaLiON/MERaLiON-GR-v1"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).to(device)
model.eval()

# 2. Make sure your waveformes are 16000 sampling rate
wav, sr = sf.read("sample.wav")
if wav.ndim > 1: wav = wav.mean(axis=1)
if sr != 16000:
    wav = torchaudio.functional.resample(torch.tensor(wav).unsqueeze(0), sr, 16000).squeeze(0).numpy()

inputs = feature_extractor( 
        [wav], sampling_rate=16000, return_tensors="pt", padding=True
    )
inputs = {k: v.to(device) for k, v in inputs.items()}

# inference. The model returns (hidden_states, logits)
with torch.no_grad():
    _, logits = model(**inputs)
        
probs = F.softmax(logits, dim=-1)
pred_class = torch.argmax(probs, dim=-1).item()
    
labels_map = {0: "female", 1: "male"}
print("Predicted gender: {labels_map[pred_class]}")

Citation

If you use this model, please cite the corresponding technical report:

@article{meralion2026gr,
  title={MERaLiON-GR: Speech Gender Recognition Model for English and SEA Languages},
  author={{MERaLiON Team}},
  journal={arXiv preprint arXiv:2608.04433},
  year={2026}
}

Related work referenced in this report:

@inproceedings{wang2025benchmarking,
  title={Benchmarking Contextual and Paralinguistic Reasoning in Speech-LLMs: A Case Study with In-the-Wild Data},
  author={Wang, Qiongqiong and Sailor, Hardik Bhupendra and Liu, Tianchi and Zhang, Wenyu and Huzaifah, Muhammad and Lertcheva, Nattadaporn and Sun, Shuo and Chen, Nancy F and Wu, Jinyang and Aw, AiTi},
  booktitle={Findings of EMNLP 2025},
  year={2025}
}

@inproceedings{cpqa_interspeech,
  title={Contextual Paralinguistic Data Creation for Multi-Modal Speech-LLM: Data Condensation and Spoken {QA} Generation},
  author={Wang, Qiongqiong and Sailor, Hardik B and Liu, Tianchi and Aw, Ai Ti},
  booktitle={Proc. Interspeech},
  year={2025}
}

@inproceedings{cpqa_asru,
  title={Incorporating Contextual Paralinguistic Understanding in Large Speech-Language Models},
  author={Wang, Qiongqiong and Sailor, Hardik B and Wong, Jeremy H. M. and Liu, Tianchi and Sun, Shuo and Zhang, Wenyu and Huzaifah, Muhammad and Chen, Nancy and Aw, Ai Ti},
  booktitle={Proc. IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  year={2025}
}