I'm curious what "word error rate" means in this context. Are they claiming that 2% of the time the model literally says the wrong word? That does not sound imp

I'm curious what "word error rate" means in this context. Are they claiming that 2% of the time the model literally says the wrong word? That does not sound impressive even compared to the TTS software of the early 2000s.