Wispr Flow is a weird startup
Wispr Flow is a weird startup right now. It is going head-on against the mammoths: Google, Muse, and the other big companies. Their voice-to-text models are at least as good, and probably better. (I will write more about the mammoth idea another time - it is about why the words we pick matter.)
NVIDIA has a leading open-source model at the moment, Parakeet (Parakeet V3), and it is multilingual. It runs on my 8 GB Mac mini without killing the other processes. Voice to text takes a few seconds, which is fine. It could be better, but it works. There are a couple of open-source models that are already very good for local work. Google has an API. OpenAI has a low-latency voice-to-text API that costs a little - it is cheap, but it is not free. That is the scene right now.
I used an open-source model for a while. The upsides: it is free, and I do not have to think about anything. The downsides: it uses memory on my machine because it runs locally, and it is a little slow.
Wispr Flow is paid, and you pay for convenience. The convenience is real. It works across all my devices, there is no word cap, it is instantaneous, and it makes very few mistakes. I do not know if it just handles my Indian accent, or if everyone is this happy with it, but I am happier with it than with the local open-source model. It works on my phone and on any of my laptops.
My one complaint is memory. My 8 GB Mac mini goes from about 60% to 70% memory usage the moment I start Wispr Flow. Not exact numbers, but it is heavy.
Even though I pay, I get a lot back. I move faster, and I have stopped typing across all my devices for days now - I just voice type. At $15 a month the cost is insignificant for what I get, and I am happy to pay it.
The weird part, and the part I do not understand yet, is why I do not have the desire to build a Wispr Flow clone. Something that works for free for most users, or costs very little, and works just as well. I am not prioritizing it, but I am surprised the market is not doing it either. There is no equivalent app in the App Store, nothing that works across Android, Apple, and everything else, and works as well. I do not get why. It is a weird setup, and honestly they are brave. They are doing it well, and as a product it works.
I do not know how this goes in the future, because AI is a fast-changing landscape.
One more thing I noticed recently: the Notetaker app inside Wispr Flow. It listens to system audio, and I did not have to configure anything. I was surprised they could listen to system audio inside a Mac - I thought that was a blocker. It takes notes in a meeting so accurately that it just works.