MERaLiON-GR: New Speech Gender Recognition Model Outperforms SOTA on English and SEA Languages

A new speech gender recognition system named MERaLiON-GR has been developed, capable of classifying voices as either female or male in English and Southeast Asian (SEA) languages. This model utilizes MERaLiON-SpeechEncoder-2, a large transformer pre-trained on an extensive speech dataset, which has been fine-tuned using Low-Rank Adaptation (LoRA) for efficiency. It incorporates a multi-scale ECAPA-TDNN network featuring attention pooling and a streamlined linear classifier. Comprehensive tests across various languages in Singapore and Southeast Asia—including English, Chinese, Malay, Tamil, Thai, Vietnamese, Indonesian, and Khmer—demonstrate that MERaLiON-GR outperforms the leading gender recognition model Vox-Profile and a significant Audio-LLM in both full-utterance and segment-level assessments. These findings highlight the importance of specialized speech models for achieving high accuracy. The research paper is accessible on arXiv with the identifier 2608.04433.