You'd have to human-label that speech. That already won't scale very well, and requires per-language annotation. Understanding and tweaking it should be done w

You'd have to human-label that speech. That already won't scale very well, and requires per-language annotation. Understanding and tweaking it should be done with hyper parameters, not semantic libraries.