She adds that text-based LLMs depend on an initial transcription step, and any error at this stage can ripple through the system — causing the chatbot to misinterpret the user’s intent and respond inappropriately.
Aw’s team at A*STAR I²R has developed MERaLiON (Multimodal Empathetic Reasoning and Learning in One Network) to address that challenge.
MERaLiON is an empathetic, culturally attuned AI model designed to support more natural and human-centric interactions across sectors, reflecting Singapore’s ambition to lead in human-centric, regionally grounded AI. It can understand and interpret the diverse speech patterns and styles of spoken Southeast Asian languages, giving it an edge over many existing LLMs, which are often trained on more Western-centric data.
According to Aw, MERaLiON builds on open-source models that have been carefully fine-tuned. “We fused OpenAI’s Whisper and Google’s Gemma before fine-tuning it on a set of audio data. The audio data comes from the National Speech Corpus, which we’ve cleaned and verified, along with insights from the research we’ve done on speech and language,” she says. Spearheaded by IMDA, the National Speech Corpus aims to be the source of open speech data for automatic speech recognition research and speech-related applications.
See also: Apple, Google warned by Germany that DeepSeek is illegal
Training the MERaLiON model required significant computing power, which was provided by the National Supercomputing Centre Singapore (NSCC) via its ASPIRE 2A+ supercomputer.
Since its launch in December 2024, MERaLiON version 1 has been downloaded over 90,000 times by start-ups, corporate research labs, media service providers, and academics.
Announced at the ATxSummit 2025 earlier today, version 2 of the LLM includes new features that improve contextual understanding and user experience.
See also: CoreWeave CEO joins world’s richest after AI firm’s 300% surge
In addition to English, Mandarin and Singlish, MERaLiON version 2 now supports Malay, Tamil, Bahasa Indonesia, Thai and Vietnamese. Support for Chinese dialects may follow in future updates. It can also handle language mixing—such as English and Chinese—a common feature of everyday speech in Southeast Asia.
What sets the enhanced MERaLiON apart is its emotional intelligence, says Aw. It can detect gender and paralinguistic features in speech, such as tone, pitch and volume, allowing for more empathetic and contextually aware interactions.
To make real-world impact
To accelerate adoption, IMDA and A*STAR I²R have launched the MERaLiON Consortium, a collaborative platform that brings together local and global industry players and R&D institutions. Members include HTX, MOH Office for Healthcare Transformation (MOHT), NCS, NSCC, SPH Media, ST Engineering, Axiom IT Solutions, BytePlus, CommonTown, DBS Bank, Grab and Microsoft Singapore.
The consortium will focus on developing practical AI applications, from multilingual customer support to health and emotional insight detection and agentic decision-making systems.
Axiom IT Solutions, for example, has developed a system powered by MERaLiON that could support social service agencies in checking in on elderly individuals living alone. The system can automatically place courtesy calls and use MERaLiON to deliver empathetic, real-time responses in the recipient’s preferred language while analysing emotional cues throughout the conversation. If signs of distress or concern are detected, the system can immediately alert caregivers for timely intervention.
Lam Pang Ngean, Axiom IT Solutions’ director for Business Development, shares that it only took “a couple of months” to develop the system. The system also features audio call logs and summaries, allowing human staff to review conversations and ensure the AI responds appropriately.
To stay ahead of the latest tech trends, click here for DigitalEdge Section
He adds: “MERaLiON has also been incorporated into our SCAM Call Protect solution, designed to safeguard the elderly from fraudulent calls. Collaborations with social service agencies are underway to bring these enhancements to life, unlocking boundless possibilities for impactful innovation.”
As for DBS Bank, it plans to experiment with MERaLiON in use cases from personalised financial and product recommendations to seamless customer servicing across languages and cultures. “Understanding local nuance, context and culture isn’t just a technical challenge, it’s a human one. DBS is delighted to be the only bank to join this pioneering effort in building and testing out a large language model that truly reflects our local voice – because innovation must speak the language of the people it serves," says the bank's group head of innovation Bidyut Dumra.
Organisations can tap into fine-tuned MERaLiON models, co-develop applications in collaboration with consortium members, or build their own solutions independently using the open-sourced base model.
MERaLiON is one of two national LLMs under Singapore’s $70M National Multimodal Large Language Model Programme, supported by the National Research Foundation and the Infocomm Media Development Authority.
A*STAR I²R and IMDA will continue to drive the growth of MERaLiON across Southeast Asia. This includes supporting early adoption use cases, developing enabling tools and infrastructure, creating case studies to showcase real-world impact, and expanding the consortium to include new members from the region.
