🔍 Project Objective Develop an AI-based Grammar Scoring Engine for spoken English audio samples.
Predict a continuous grammar score between 0 and 5 for each audio file.
🎧 Audio Transcription Used OpenAI Whisper (base) model to convert audio to text.
Ensures accurate, real-time transcription of 45–60 second clips.
🧠 Linguistic Feature Extraction Extracted features using:
spaCy: sentence and token analysis, POS tagging.
language-tool-python: grammar error detection.
textstat: readability metrics (Flesch, Kincaid scores).
NLTK Sentiment Analyzer: sentiment score.
Computed metrics like:
Number of sentences, grammar errors, token length
POS ratios (noun, verb, adjective)
Lexical diversity, sentiment score, and more.
🔡 Text Embeddings Generated semantic embeddings using Sentence-Transformers (all-MiniLM-L6-v2).
Captures deeper linguistic patterns and context in the spoken content.
🏗️ Feature Engineering Pipeline Combined:
Linguistic features (14+ metrics)
384-dimensional sentence embeddings
Created a complete feature set for modeling.
🤖 Model Training & Evaluation Trained multiple regression models:
Ridge, Random Forest, Gradient Boosting, XGBoost
Split data into training and validation sets.
Evaluated using:
Pearson correlation
RMSE and MAE metrics
Visualized prediction performance with scatter plots and regression lines.
🔍 Feature Importance Analyzed top contributing features using feature_importances_ from tree models.
Helped interpret model decision-making.
📦 Final Predictions & Submission Processed test audio files using the same pipeline.
Generated grammar scores and clipped to the [0, 5] range.
Saved results in submission.csv for Kaggle evaluation.