Apple runs on-device on older models, too, just wimpier models.

human resources (my voice and time) are far more valuable than the system resources. going to the cloud is absolutely worth it to prevent a typo

That doesn't work if you have limited or no connectivity (e.g. on a mountain range). There are also privacy concerns, e.g. a doctor using it to transcribe medical information.

Apple's stt has been on-device for a long time now, long before iPhone 16. I haven't noticed any improvements since my first ever iphone 5S. I'm pretty sure wispr flow can use on-device models. I use Voiceink[0] which can use parakeet models on-device and can optionally use cloud models.It's like night and day comparing Apple's to Voiceink. The only advantage I find to Apple's stt is less friction. 3rd party apps just can't integrate as smoothly with the system. There's a gesture to activate Appledictation when Voiceover is on.

It's been around and available as an API to devs since at least 2021 in iOS. The problem was even on the best iPhone at that time, I could never get it past ~0.8x speed and after 15-20 minutes the device would heat up so much the display dimmed.

For context, I was working on a podcast app with on-device transcription, had to park that idea for years before it got to today's performance.