What is balanced accuracy?
Balanced accuracy is the average of a classifier's true positive rate (sensitivity) and true negative rate (specificity). It measures how well a model performs on each class independently, rather than letting the majority class dominate the score.
What is an example of balanced accuracy?
On a dataset where 95% of frames are non-speech and only 5% are speech, a model that always predicts "non-speech" scores 95% on plain accuracy but is useless. Balanced accuracy correctly reports it at 50% - because its sensitivity is 0% and its specificity is 100%, averaged.
How does balanced accuracy work?
Balanced accuracy = (sensitivity + specificity) / 2. It assumes both classes are equally important, regardless of how often they appear. This makes it a natural fit for imbalanced datasets, where standard accuracy is misleading because guessing the majority class already gets a high score.
How does ai-coustics use balanced accuracy?
Balanced accuracy is useful for evaluating detection models like Quail VAD on real voice traffic, where speech and non-speech frames are rarely evenly distributed. Reporting balanced accuracy alongside F1 gives a fairer picture of how a detector behaves on both classes, which is what matters when the cost of missing a barge-in is very different from the cost of a false trigger.
