Mozilla Common Voice has a dataset with hundreds of hours of recorded Cantonese sentences: https://mozilladatacollective.com/datasets/cmu5x3i5w00d6o107... Hardly scarce enough to require AI audio for listening practice.