I’ve been following recent progress in speech AI, and one thing I’ve been wondering about is whether current limitations are increasingly caused by training data rather than model architecture. Models seem much better than they were a few years ago, ye…