This is motivated by some feedback from Andreas about how it is currently hard for users to figure out how to manually set the threshold values in the Parrot post-processor (and I agree with him).
Dhruv Jain 2022's system ("Nonverbal Sound Detection for Disordered Speech") sets the threshold values in their post-processor by finding the threshold values that optimize certain metrics, on a per-class basis:
This would be simple to implement for Parrot once we implement / expose the metrics we're interested in, and I definitely want to try implementing this.
This is motivated by some feedback from Andreas about how it is currently hard for users to figure out how to manually set the threshold values in the Parrot post-processor (and I agree with him).
Dhruv Jain 2022's system ("Nonverbal Sound Detection for Disordered Speech") sets the threshold values in their post-processor by finding the threshold values that optimize certain metrics, on a per-class basis:
This would be simple to implement for Parrot once we implement / expose the metrics we're interested in, and I definitely want to try implementing this.