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
About the Project
Can artificial intelligence tell whether you are a good singer?
Inspired by our shared interest in music, we built an AI-powered singing scorer that predicts the quality of a singing performance on a scale from 1 to 5.
Training the Model
We trained our machine learning system on singing data using a MERT audio model with an MLP regression head to predict a Mean Opinion Score (MOS), considering common factors such as pitch and stability.
What moved the needle for us was sourcing higher-quality training datasets. This improved the model’s performance and helped us achieve a prediction correlation of over 0.7, where 1 represents a perfect correlation.
From Model to Web Application
Our web application allows users to record their voice directly or upload an audio clip and instantly receive a predicted score, accompanied by funny meme reactions.
Watching our batchmates test their voices and laugh at the results became one of the highlights of the project, turning the showcase into an entertaining community experience.
Our Project Journey
Like many machine learning projects, our journey in the 11th Data Associate Programme was not without challenges.
We faced discouraging periods when model performance stagnated despite numerous experiments, schoolwork became stressful, and we even ran out of GPU credits during training.
With support from our mentors, we were eventually able to reach our targeted milestones, improve the model, and successfully deliver a working application.
Project Satisfaction
We felt most satisfied with creating an AI application that was genuinely fun to use. Seeing our batchmates test their singing and laugh at the results made the project especially rewarding.
Key Takeaway
The experience reinforced for us that building successful machine learning systems requires more than selecting the right model. It also requires high-quality data and an experience that encourages people to interact with the technology.