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, yet they still struggle with regional accents, code-switching, spontaneous speech, and speakers who don’t match “standard” pronunciation.
My guess is that collecting this kind of data at scale is much harder than collecting carefully scripted recordings.
If you were building a speech model today, where would you invest more effort: better models or more diverse speech data? Why?
[link] [comments]