Meet Lemka – The Speech AI Built for Khasi, Garo, Mizo, Meitei, Nagamese and Kokborok

For millions of people across Northeast India, using digital services has long meant switching to English or Hindi. Welfare schemes, government helplines, farming advice, and online learning tools rarely understand the languages spoken at home. Regional voices have largely been left out of the conversation.

That is beginning to change.

Shillong-based research and AI deployment lab MWire Labs has released Lemka, an end-to-end Speech AI system designed specifically for six Northeast Indian languages: Khasi, Garo, Mizo, Meitei (Manipuri), Nagamese, and Kokborok.

Lemka allows a person to speak in their own language and be understood, and for the system to reply in the same language — without typing, without reading, and without needing English or Hindi. In simple terms, it brings practical voice AI capability to languages that mainstream systems still struggle with.

How Lemka Works

Native voice input (Khasi, Garo, Mizo, Meitei, Nagamese, Kokborok) ↓ Lemka STT (Speech-to-Text) – trained on 1,000+ hours of real field audio ↓ Text / AI response ↓ Lemka TTS (Text-to-Speech) – natural voice output ↓ Spoken reply in the user’s language

The entire pipeline is built for real conditions in the region, including offline and edge deployment.

Why Lemka Is Emerging as the Go-To Speech AI for These Languages

Most systems are trained mainly on web data and perform poorly on the actual spoken forms of Northeast languages. Lemka takes a different path.

• Built on real field audio: More than 1,000 hours of proprietary recordings collected across urban and rural areas of the region.
• True voice-in, voice-out: Combines high-quality speech-to-text (Lemka STT) and natural text-to-speech (Lemka TTS), enabling real voice AI interactions.
• Offline and edge-ready: Designed to work without constant internet — critical for remote health centres, farms, and areas with weak connectivity.
• Privacy-first architecture: Supports a Bring Your Own Cloud (BYOC) model, allowing government departments and organisations to keep sensitive data on local servers.

These strengths make Lemka one of the most practical Speech AI systems currently available for Khasi, Garo, Mizo, Meitei, Nagamese and Kokborok.

Real Impact Across Northeast India

Lemka builds on earlier work by MWire Labs, including the open release of NE-ASR (a speech recognition model covering eight Northeast languages on Hugging Face) and an offline voice assistant developed for Garo-speaking visually impaired users.

By making voice the main interface, Lemka opens clear possibilities:

• Government services: Citizens can call helplines or check welfare schemes in their native language.
• Agriculture: Farmers can receive spoken advisories on their phones while working in the field.
• Education: Students can interact with learning tools that speak their primary language, reducing dependence on literacy in English or Hindi.

A Clear Purpose

“Our languages carry our history, culture, and identity,” says Badal Nyalang, Director of MWire Labs. “With Lemka, we’re not just building speech models — we’re making sure Northeast India’s voices are actively woven into the future of technology, rather than left behind.”

Lemka forms part of the larger NE-Stack, MWire Labs’ effort to build foundational language AI for the region’s more than 220 indigenous languages. Most of these languages still have little or no AI support.

For communities that have waited years to hear their languages properly recognised by machines, Lemka is a meaningful step forward. It is not a generic model with a few regional languages added. It is a Speech AI system built from the ground up for the linguistic reality of Northeast India — and one of the strongest practical voice AI options available for these six languages today.