Voice agents in a four-language city

Singapore has four official languages: English, Malay, Mandarin and Tamil. A voice agent built for one of them will meet customers who use two or three in the same call.

The latest household survey shows how the mix is moving. In 2025, 58.1% of residents aged five and over spoke English most often at home, up from 48.3% in 2020, while Mandarin fell from 29.9% to 26.6%.¹ English is the majority home language now, and many of those households also speak another language at home.¹

Local models exist

A*STAR’s Institute for Infocomm Research built MERaLiON-AudioLLM for this setting, describing it as the first general-purpose audio language model designed to understand Singlish, fine-tuned on 260,000 hours of speech and audio data.² Its successor, MERaLiON-2, adds Malay, Tamil and other Southeast Asian languages, and its developers report leading results on Singlish, Mandarin, Malay and Tamil speech recognition.³ Those are the developers’ own benchmarks. Your callers are the benchmark that counts.

Test on real calls

A demo in clean English says little about a call centre line in Singapore. Build the test set from recorded calls, with the share of each language and of mixed-language calls matching your traffic. Measure recognition accuracy and resolution by language segment, so a strong average cannot hide a weak segment.

Hand over in the customer’s language

87% of customers say a company using GenAI for service must offer a route to a human agent.⁴ In a multilingual market, that route should reach a person who speaks the language the customer was using, with the transcript and the reason for the handoff already in front of them.