README.md

metadata

title: Indic Sentiment Audio App
emoji: 📚
colorFrom: yellow
colorTo: green
sdk: gradio
sdk_version: 6.4.0
app_file: app.py
pinned: false
license: mit
short_description: Real-time audio sentiment analysis for code-mixed IndicLang

🇮🇳 Project-IV: Real-Time Indic Sentiment Analysis

Audio-Visual Sentiment Analysis for Code-Mixed Indian Languages (Gujlish & Hinglish)

🚀 Project Overview

This application represents the final deliverable for Project-IV. It is a sophisticated AI system designed to solve a specific challenge in Natural Language Processing (NLP): Sentiment Analysis of Code-Mixed Indian Languages.

Standard AI models fail when users mix languages (e.g., "Aa movie bahu saras che but ending weak hatu"). This project solves that problem using a custom fine-tuned architecture.

🛠️ Technical Architecture (The Pipeline)

This application uses a Two-Stage Pipeline to process real-time audio:

1. Stage 1: The Ears (Automatic Speech Recognition)

   • Model: openai/whisper-small
   • Function: Captures live audio from the microphone and transcribes it into text. It is robust enough to handle Indian accents and mixed-language speech patterns automatically.

2. Stage 2: The Brain (Sentiment Classification)

   • Model: marshal-yash/gujlish-sentiment-analysis
   • Architecture: Google MuRIL (Multilingual Representations for Indian Languages).
   • Training: Fine-tuned on a proprietary synthetic dataset of 150,000 samples (gujlish_150k_massive.csv).
   • Performance: Achieved 100% Accuracy on the validation set during training.

📊 Dataset Details

The "Brain" of this system was trained on a massive, diverse dataset generated specifically for this project:

• Size: 150,000 unique data points.
• Languages: Gujarati-English (Gujlish), Hindi-English (Hinglish), and Pure English.
• Domains: Technology, Movies, Food, Sports, and Daily Life conversations.
• Technique: Generated using advanced combinatorial data augmentation to ensure robust handling of grammar and vocabulary variations.

🎯 How to Use

1. Allow Microphone Access when prompted by the browser.
2. Click the Microphone Icon to start recording.
3. Speak a sentence in Gujarati, Hindi, or English (or a mix of all three!).
   • Example: "Server connect nathi thatu, bahu slow che."
   • Example: "Wow, what a performance! Maja padi gai."
4. Click Stop Recording and then Submit.
5. View the transcribed text and the AI's sentiment prediction (Positive/Negative/Neutral).

---

Developed by: Yash Bharvada Model Hosted on: Hugging Face Hub