It is definitely doable to get openly licensed data, you just have to do it via voluntary participation of crowdsourced data acquisition programs. For example the RNNoise model was retrained from such crowdsourced data.