Spoken Grammar Scoring Engine

A multimodal machine learning model for scoring spoken English grammar on a continuous 0–5 scale.

The model combines:

• Acoustic speech features
• Whisper-based speech transcription
• Linguistic features from the transcript
• MiniLM semantic embeddings
• Ensemble regression

Performance

• 5-Fold CV RMSE: 0.6876
• 5-Fold CV Pearson: 0.8338
• Kaggle Public Score: 0.5278

The model was developed for the SHL Hiring Assessment 2026 and is an independent implementation for spoken grammar assessment.

Author: Satyam Kumar