Shabdasakhi (शब्दसखी)

Tagline: Say it in richer Marathi. / तुमची मराठी, आणखी सुंदर

Short Description

Shabdasakhi is an open-source Marathi expression coach. Speak or type a sentence, even one mixed with Hindi or English, and get two improved Marathi versions: Natural & Polite and Formal & Scholarly. Learn from word upgrades you can save. Powered by Gemma running locally.

Full Description

Many Marathi speakers naturally mix Hindi and English words into everyday conversations. Shabdasakhi helps people who already speak Marathi discover richer, more natural Marathi expressions without judging how they speak.

Users can speak or type a sentence and receive two rewritten versions: Natural & Polite for everyday communication and Formal & Scholarly for more formal expression.

Whisper, powered by faster-whisper, converts speech into editable text locally. Gemma 4 (E4B), running locally through Ollama, rewrites the sentence in both styles and explains individual word upgrades. Users can save useful words to their personal vocabulary list.

A community-maintained glossary guides Gemma toward preferred Marathi vocabulary. Anyone can contribute new words and improvements through GitHub pull requests.

Shabdasakhi is designed around local processing: models are downloaded once, and inference runs on the user's machine without relying on a cloud AI service.

The project is open source, and community contributions can help improve Marathi language quality, vocabulary coverage, and reliability.

How Gemma Is Used

Gemma 4 (E4B) powers the core language transformation. It interprets code-mixed Marathi, Hindi, and English sentences, rewrites them in two Marathi styles, and explains the word choices and upgrades. A local glossary provides additional vocabulary guidance.

Built With

Gemma 4 (E4B), Ollama, faster-whisper, Streamlit, Python, PyYAML.

Known Limitations

Marathi quality depends on the capabilities of the local model, and generated expressions may be inaccurate or unnatural. Users should review outputs before relying on them. Marathi speech recognition is experimental and may produce transcription errors.

What's Next

• Marathi voice output.
• A daily Marathi word and expression.
• Support for additional regional languages.
• Improved Marathi quality through community feedback and contributions.

Track: Best Projects with Gemma

Repository: https://github.com/samarth-murade/shabdasakhi

Demo Video: https://youtu.be/LQicxFUJOhw