We need to separate three separate issues. I work with call centers and I always need to discuss all three

1. Voice to text transcriptions

2. Text to understanding

3. Adding capabilities where it can do something with #2.

The voice to text that Siri uses seems to be worse than when you are dictating using voice to text from the keyboard.

The latter gets close to 100% with my southern native English accent and does okay when I’m trying to speak Spanish. Siri messes up with English a lot more and it’s a lost cause when I try to speak Spanish.

I have a British accent and the speech-to-text from the keyboard is also terrible.

Honestly with these assistants I'd rather just type my query. Voice input is embarrassing and error-prone. The only place that voice input is useful is in the kitchen.

When they switched to "ML" (this was way before ChatGPT ate the world) Siri's speech to text went to absolute shit.

I suspect they had a very carefully hand-crafted model before, and replaced it with an ML model "that will fix itself over time" - and it never did.