But Apple's uses so few system resources and runs fully on device on newer iPhone models (16+ I believe). It's so efficient. I really enjoy using Handy with Parakeet as the model, but the system resource usage is a monster compared to Apple's (although very good).
replyLooks like Wispr Flow uses a cloud model [0]:
> Cloud based speech processing infrastructure for 1B users
It gets to be a messy comparison because my iPhone can do STT with no latency pretty well fully on device, but Wispr Flow requires a cloud model, but to be fair, older Apple devices do as well. It's not an apples and oranges comparison, but I think those technical details make this a non direct comparison in a few ways.
For on-device with low system resource usage, Apple's is pretty damn good.
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.
replyIt'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.
replyFor 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.
Apple runs on-device on older models, too, just wimpier models.
replyhuman 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
replyThat 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.
replyFWIW - I also really like Wispr Flow, but I moved to running the 'Whisper Large' model locally using Handy (https://github.com/cjpais/Handy), which has been essentially as good, while also having lower latency.
replyHandy is great. It exposes a bunch of open models beyond Whipser too, and though I haven’t tried too many of them, I’ll throw in a rec for the Parakeet model which feels pretty much on par with Whisper for accuracy and is way way faster.
replyI’d say STT is pretty much a solved problem. Everyday there is a new product and can be one-shotted by any current top of the line LLMs. Take a look at this [1]. Apple is just stuck in the past.
reply