Singing Scorer
An ML model trained to evaluate singing performances using objective criteria like pitch accuracy and vocal stability — because human judgment is subjective.
Project Details
Singing Scorer is a web application that uses a fine-tuned MERT-v1-95M model with a custom regression head to score singing recordings on objective acoustic criteria such as pitch accuracy and stability. The frontend is built with React 19 and Vite, with a Flask backend handling audio processing and inference.
Human judgment of singing quality is inherently subjective and inconsistent. The team wanted to train an ML model to evaluate performances against objective acoustic metrics — pitch accuracy, vocal stability, and tonal consistency — removing the variability of human scoring. The project was also driven by curiosity and a desire to build something genuinely fun to use.
Project Satisfaction
Team Satisfaction: 4/5. The model achieved a meaningful level of accuracy in scoring singing performances. There is room to improve further, particularly in edge cases with unconventional vocal styles.
Key Takeaway
Model accuracy is an iterative process. The team identified that further fine-tuning with a more diverse training dataset would significantly improve scoring consistency across different vocal ranges and styles.