NAMMA SPEAK is an AI-powered Kannada–English code-mixed language assistant that aims to make AI understand and generate text in the way people naturally communicate.

People frequently mix Kannada and English in everyday conversations, using informal expressions, Kannada script, and Kannada words written in English letters. General-purpose language models can struggle with these variations.

Our solution explores adapting Google’s Gemma language model using parameter-efficient LoRA fine-tuning, Unsloth, and the AI4Bharat IndicCMix dataset.

NAMMA SPEAK supports eight planned conversion tasks across English, Kannada–English code-mixed text, full Kannada script, and Romanized Kannada. An interactive Gradio application lets users enter text, select a conversion direction, and view the generated output.

We evaluate the original and fine-tuned models using SacreBLEU chrF and inspect generated responses to identify errors and opportunities for improvement.

Our goal is to explore how efficient fine-tuning can improve regional-language AI while building a practical, accessible language-conversion tool.