MERaLiON
Singapore's national audio-language models for Southeast Asian speech
Overview
MERaLiON is a family of speech-text models from A*STAR, built under Singapore's S$70 million National Multimodal LLM Programme. It transcribes, translates, summarises and answers questions about speech in English, Mandarin, Malay, Tamil, Indonesian, Thai and Vietnamese, plus Singlish, Cantonese and Hokkien, and reads emotion and other cues from audio. Weights are on Hugging Face under A*STAR's own licence, and a hosted API is available.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 19 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with MERaLiON
- Transcription of Singlish, code-switched and Southeast Asian speech
- Spoken question answering and summaries taken straight from audio
- Emotion, gender and age-group cues read from speech
- Hokkien speech generated from Mandarin input (OmniVoice-Hokkien-TTS)
- Self-hosted deployment from Hugging Face weights, including MLX and GGUF builds
- An OpenAI-compatible hosted API for chat, transcription and emotion analysis
ASEAN Perspective
MERaLiON in Southeast Asia
MERaLiON is Southeast Asia's most seriously funded attempt at a speech model that understands how the region actually talks. A*STAR's Institute for Infocomm Research built it under Singapore's S$70 million National Multimodal LLM Programme with NRF and IMDA backing, and the point is accented, code-switched Southeast Asian speech — the input on which Western speech models degrade sharply. It is designed to close the performance gap for voice products aimed at users in Singapore, Malaysia, and Indonesia. Being a national research programme, it is also published openly rather than held as a commercial advantage, which makes it usable by organisations across the region rather than only in Singapore.
MERaLiON is a purpose-built option for Singapore-accented and code-switched speech: an end-to-end audio model that handles Singlish, Hokkien and Southeast Asian languages, with free weights and a hosted OpenAI-compatible API. Its 2026 releases added paralinguistic reasoning, a fast 3B speech-recognition model and Hokkien speech synthesis.
Weigh two things. The licence is A*STAR's own, carrying Gemma and Whisper terms rather than a standard open licence, and nearly all benchmarks are the team's own. Best for builders of local voice and call-centre products who will test it on their own audio.
What we found
MERaLiON (Multimodal Empathetic Reasoning and Learning in One Network) is a family of speech-and-text models built for Singapore and Southeast Asia, one of two national models under Singapore's S$70 million National Multimodal LLM Programme, backed by the National Research Foundation and IMDA. It first launched in December 2024 from A*STAR's Institute for Infocomm Research (I²R), which on 1 July 2026 merged with the Institute of High Performance Computing into the A*STAR Institute of Advanced Intelligence and Computing (IAIC). The models work on raw audio rather than transcribing first, so they can answer questions about a recording, summarise it, and read tone, emotion, age group and background sound.
Sixteen models sit on Hugging Face, from the original AudioLLM through MERaLiON-2 and the flagship MERaLiON-3-10B to a 3B speech-recognition model and a Hokkien text-to-speech model released in May 2026. The weights are free to download but not under a standard open licence: the MERaLiON Public Licence is A*STAR's own and bundles the Gemma and Whisper terms of the models it builds on. The project lists seven core languages, English, Mandarin, Malay, Tamil, Indonesian, Thai and Vietnamese, plus Singlish, Cantonese and Hokkien. The first release was accepted as an ACL 2025 system demonstration paper.
The team's own model card says MERaLiON-3-10B beat Qwen3 Omni, Gemini 3 Flash and GPT-4o Audio on 31 of 59 benchmarks in age, gender, spoken question answering and paralinguistic tasks, while claiming only to stay competitive with its predecessor on transcription. A*STAR's own API terms warn that emotion analysis is probabilistic, may be biased, and should not be the sole basis for decisions about people.
A hosted route exists too: a playground and OpenAI-compatible API at meralion.org, with transcription, chat and emotion-and-gender endpoints, powered partly by a consortium-tuned ASR model that is not among the public downloads. No API prices are published. The industry consortium reached 19 members in August 2026, with Standard Chartered and UOB the newest.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including MERaLiON's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with MERaLiON unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to MERaLiON directly →
Spotted something out of date? Suggest an update →
Alternatives to MERaLiON
More in LLMs & Chat
