Hey, naanu late aagthini!

ನಾನು ತಡವಾಗಿ ಬರುತ್ತೇನೆ

NAMMA SPEAK

Making Kannada–English chat feel natural

A fine-tuned Gemma assistant for everyday code-mixed language.

Kannada script

Romanized Kannada

English

CODEMITRA AI • KANNADA–ENGLISH CODE-MIXED LANGUAGE ASSISTANT

01

Teach gemma to understand code-mixed regional language

Code-mixed text

The problem: people don't chat in just one language

Everyday messages mix English, Kannada words, scripts, and informal spelling.

TODAY'S CHALLENGE

• Informal Romanized words can be ambiguous

• Literal translation may sound unnatural

• The requested script or style can be missed

OUR SOLUTION

• Fine-tune Gemma on real code-mixed examples

• Convert between English, Kannada script, and Romanized Kannada

• Keep the meaning and conversational tone

Goal: make language conversion fit the way people actually message.

CODEMITRA AI • KANNADA–ENGLISH CODE-MIXED LANGUAGE ASSISTANT

02

How NAMMA SPEAK works

A compact, measurable fine-tuning pipeline

01

IndicCMix

Kannada examples in four text forms

02

Prepare data

Split first; build conversion tasks

03

LoRA fine-tune

Adapt Gemma efficiently on a T4 GPU

04

Evaluate

chrF + Kannada-speaker ratings

05

Demo

Gradio app for live conversion

›

›

›

›

Fair comparison: same held-out test examples for plain, prompted, and fine-tuned Gemma.

CODEMITRA AI • KANNADA–ENGLISH CODE-MIXED LANGUAGE ASSISTANT

03

What we will demonstrate

A live website — plus evidence that the model actually improved.

LIVE DEMO

Type a message → choose a format → convert

• English

• Kannada script

• Romanized Kannada

• Code-mixed text

HOW WE MEASURE SUCCESS

• chrF score on unseen test examples

• Human ratings for naturalness and meaning

• Output-format accuracy and response speed

• Error analysis: what improved and what did not

Our promise: report measured results, not assumed improvements.

CODEMITRA AI • KANNADA–ENGLISH CODE-MIXED LANGUAGE ASSISTANT

04