That's not because voice recognition doesnt work. It's been working quite well for 15 years now. Just ask the Deaf community.
The issue is even if the error rate is 1 in 100 and you scale that up to a billion users supporting thousands of use cases, you then need to maintain a large army of people to manually handle those errors.
Its a huge cost to even the largest firms. Very similar to what Facebook/Twitter/Youtube end up doing to handle content moderation. The faster systems scale the faster you get swamped with bugs you don't have bandwidth to fix. The consequences begin to multiply. Your basic Jurassic Park story plays out.