When AI Speaks Your Language

Asia is building AI that talks in local languages - so more people can use it without help, especially older adults, non-English speakers, and users outside big cities.

Lost in the 250-million-strong tide at India’s Maha Kumbh? A Bhashini chatbot can now guide you to the nearest ghats, in Hindi, Tamil, or Bengali. Selling fried snacks on a street in Indonesia? GoPay’s voice assistant understands you perfectly.

Across Asia, technology is finally starting to speak like its people do.

And the shift goes beyond adding more languages. Asia is starting to build culturally grounded AI systems that challenge the English-first design of the early internet.

This matters because the region is home to more than 2,300 languages, and yet for decades most digital platforms defaulted to English. UN data shows the region’s digital gaps persist in rural and ageing populations - groups often excluded when technology does not speak in their mother tongue.

That’s changing quickly. A wave of local-language AI systems is reshaping how people access information, payments, public services, and identity verification. Voice recognition is more accurate. On-device AI is cheaper. And multilingual models are being embedded directly into the apps people already use.

“Localizing both language and cultural behavior ensures that AI interacts naturally and respectfully, becoming a tool that feels like it ‘belongs’ to the community rather than being ‘imposed’ from outside.”

Wenxuan Zhang, SUTD Assistant Professor, Information Systems Technology and Design Pillar, Singapore University of Technology and Design

This trend reflects something deeper. Asia’s next leap in AI won’t be measured in model size, but in whether technology can think, speak, and respond like the people it serves. Local-language AI is becoming essential infrastructure and a quiet challenge to the English-first default that shaped the early internet.

Local-Language AI in Action

Across the region, this evolution is taking concrete shape. Four countries in particular, India, Indonesia, Malaysia and Singapore, illustrate different pathways.

India’s Mission Bhashini is slowly becoming the country’s linguistic backbone for AI. Recognised by UNICEF, it enables apps and chatbots to operate in 22 languages via open APIs. It powers more than 350 AI models serving ministries and state governments, and has processed over a billion interactions/queries - from WhatsApp assistants helping citizens query government schemes, to voice bots that process KYC-lite verification or digital payments for users not comfortable with English.

Yet the digital imbalance remains stark: English accounts for roughly 50% of all web content, while Indian languages add up to less than 0.1 percent. To change that, Bhashini is growing. New projects like Adi Vaani are digitising endangered tribal languages like Gondi, Bhili, and Santali. India’s approach shows what state-led, open infrastructure can achieve at national scale.

Indonesia by contrast, is leaning on commercial ecosystems and sovereign cloud strategy to align language technology with data security. Sahabat-AI is taking multilingual AI mainstream. Built by GoTo Group and Indosat Ooredoo Hutchison and running on GPU Merdeka, a sovereign cloud that keeps data inside Indonesia. It’s powered by a 70-billion-parameter model that understands Bahasa, Javanese, Sundanese, Balinese, and Bataknese.

Its open-source versions have been downloaded more than 35,000 times, enabling use cases in education, customer support, and public services. For ordinary users, it’s already powering natural voice interactions in GoPay and on sahabat-ai.com. GoPay itself serves more than 60 million annual transacting users, and the voice assistant handles millions of spoken interactions every month, showing how quickly language AI can scale when embedded into daily transactions.

Where Indonesia is optimizing for commercial scale and sovereign-cloud control, Malaysia is betting on something different: identity, law, and cultural context.

Malaysia’s bet is ILMU (Intelek Luhur Malaysia Untukmu), a large language model developed entirely in-country by YTL AI Labs. A standout strength of ILMU is its understanding of local context. For example, when asked about medical marijuana, ILMU’s responses reflect Malaysia’s current legal stance - prohibitive, with narrow medical exceptions.

Early data shows ILMU outperforming global models in Malay-language tasks. It has already partnered with over ten Malaysian companies, from Carsome to RHB Bank, to bring local-context AI to everyday products.

Singapore’s strategy is outward-looking: to position itself as Southeast Asia’s multilingual AI hub. Its SEA-LION model or the Southeast Asian Languages In One Network, is a cornerstone of that plan. Developed by AI Singapore under a S$70-million national initiative, SEA-LION supports more than 11 regional languages, from Malay and Thai to Vietnamese, Tagalog and Tamil.

Open-source by design, it’s already being tested by banks for cross-border customer service, law firms reviewing multilingual documents, and by schools building translation and learning tools in students’ native languages.

Hard Problems

But multilingual AI comes with its own set of issues. Dialects and accents vary every few kilometers, making speech recognition far harder than in English. Tamil in Jaffna isn’t Tamil in Chennai. Javanese changes drastically between Central and East Java.

As SUTD’s Wenxuan Zhang notes, the challenge isn’t just linguistic. It’s structural. Asian languages often lack the clean, labelled data that English models take for granted.

“Existing open-source resources are often scarce, messy, or unevenly distributed across different languages and dialects. For example, Malay and Indonesian are closely related yet differ in subtle lexical and cultural ways, so models trained on one often misinterpret the other. Code-mixed languages like Singlish bring an additional layer of complexity.”

These variations mean models must decode mixed grammar, tone, and nuance without losing meaning. Data scarcity makes this even harder: many Asian languages lack large, clean datasets, and their scripts often have inconsistent spelling or orthographic rules.

Then there are questions of ownership: who owns the voices and text used to train these models? Community radio, regional news, and social media often sit in grey legal and ethical zones.

Reliability is another challenge. In areas like healthcare or finance, a mistranslated instruction can have consequences. That’s why developers are turning to RAG or retrieval-augmented generation, citations, and human fallback systems to keep models grounded.

Values, Politics & Ideology

Those same concerns of ownership, bias, and reliability, take on a sharper edge as adoption accelerates. Language is never neutral. AI doesn’t simply translate, it carries the values and assumptions baked into its training data. What seems neutral in one culture can sound political in another. Developers need to reflect local norms without slipping into state narratives, especially on issues like religion or elections.

As Zhang notes, multilingual AI reduces some Western-centric bias but doesn’t eliminate it. The underlying data balance still skews heavily toward English.

“English still tends to dominate pretraining/alignment data. The challenge is not only removing bias, but ensuring that models adapt fairly to each cultural context without reinforcing harmful local prejudices or flattening cultural diversity into a single global norm.”

These concerns make accountability even more critical. When a model mistranslates medical advice or financial instructions, responsibility becomes murky: is it the developer, the data provider, or the agency deploying it?

Most Asian countries now have data-protection or AI-ethics frameworks, but few outline what good governance looks like across dozens of languages and dialects.

Researchers are proposing:

• Language-specific safety cards listing data sources, known limitations, and performance by language.

• Incident-reporting channels where users can flag harm in their own language.

• Culturally aware evaluation tests instead of English-centric benchmarks.

But even these safeguards don’t fully address the deeper question of purpose. That’s where voices like Elina Noor, Senior Fellow at the Carnegie Endowment for International Peace, come in. She warns that once AI speaks everyone’s language, its potential impact widens dramatically. These models can support public services and inclusion, but they can just as easily be used for monitoring citizens, shaping narratives, or extracting data from communities that have historically had little protection.

“Local language models, assuming they are designed, developed, and deployed with care and conscience, could go a long way towards supporting communities in crucial services like education.”

But she emphasizes that capability isn’t the core issue.

“The big caveat is what applications are these models going to be used for and whether those benefits are going to outweigh the models’ true costs (ecological, labor, etc.) across their entire life cycle.”

Her caution extends to the everyday rush to automate.

“Instead of assuming that having chatbots for everything is a good thing, we should question whether it’s even a good idea in the first place. Just because it can be done doesn’t mean it should be done. This should be the original safeguard.”

Asia stands at an inflection point. The region is proving that these systems aren’t simply catching up to English models; they’re redefining what “inclusive AI” should mean in a multilingual world. And in doing so, Asia is showing that the future of AI can be broader, fairer, and built for everyone.

For More Info on Asia Tech Lens