🎙️ A voice-based ML app that predicts grammar fluency score (0–5) from spoken audio samples using audio feature extraction (MFCC) and regression modeling.
Built with Gradio, deployed on Hugging Face Spaces, and trained with real audio samples.
Try the App on Hugging Face Spaces
No installation required – works directly in browser
The Grammar Scoring Engine is a machine learning project that leverages audio signal processing and regression modeling to evaluate grammar fluency in spoken English.
Users are prompted to speak naturally for 45–60 seconds, and the model provides an objective grammar score between 0 and 5, based on audio features.
- Voice recording using Gradio UI
- MFCC feature extraction via
librosa - Regression model trained on audio-annotated dataset
- Evaluation using Pearson correlation
- Deployed on Hugging Face Spaces (free, public, portable)
| Component | Tool/Library |
|---|---|
| UI & Deployment | Gradio + Hugging Face Spaces |
| Audio Processing | Librosa |
| ML Model | Scikit-learn (LinearRegression) |
| Backend | Python |
| Packaging | Joblib |
- Dataset: 444 training samples, 195 test samples (45–60s voice recordings)
- Target Variable: Continuous grammar score [0, 5]
- Preprocessing:
- Noise handling
- Silence detection
- Resampling to 16kHz
- Feature Engineering:
- MFCCs (13-coefficients)
- Feature selection with correlation thresholding
- Model: Linear Regression
- Evaluation Metric: Pearson Correlation Coefficient
Made with 🧡 by Manmath Balaji Hatte