Do yall use dictation when talking to AI?

@hosseinyazdi I see, so accuracy and what field you're currently talking about matters more than words per second. That's honestly something I haven't thought about, a fast switch so it's coding mode, technical professional mode, or just normal conversation(refining features, brainstorming etc...). The app currently has a dictionary for any user custom words, the downside is that the dictionary is a flat, always-on list you have to maintain by hand: it never learns from the words you fix, it can't tell that "class" means one thing in code and another in conversation, and once it grows past a few dozen entries whisper quietly ignores the ones at the end.

This could genuinely be something to work on and something that will make this app a lot better, cause most dictation apps don't have this feature of what language(field) the user is currently talking in. I will note this down.

As for accuracy of words, aside from technical terms, whisper large turbo V3 has actually been fast and accurate, also long transcriptions almost never fail, and if they do, the app already has your recordings saved to transcribe again, you don't have to say the whole thing again.

Fr these are some valuable advice and I want to offer you lifetime access of phantom voice, is there anyway I can contact you and send you the phantom ID to activate lifetime access? or you can just email me here leftinverse@gmail.com I would love to keep in touch with you and better this app. Thanks bro

@hosseinyazdi I built this after your comment, then took it out again, so I owe you the reasoning. The scoping part was fine. What killed it was the switch itself. For modes to help, switching has to be as cheap as holding the dictation key - anything that costs a trip to the menu bar just doesn't happen mid-flow. And a mode you forget to switch isn't neutral: you get the wrong vocabulary biasing recognition toward words you didn't mean. That's worse than one flat list, which at least fails the same way every time. It's also invisible state. You can't tell which mode you're in without going to look, so being in the wrong one stays undetectable until the transcript comes out wrong. Inferring the mode automatically has the same problem with less control. So the real cost wasn't per-mode vocabularies, it was a hotkey plus a permanent on-screen indicator of which mode you're in - and at that point I'm adding modality to an app whose entire pitch is hold a key and speak. I'd rather not ship a feature whose failure is silent and worse than not having it. What did come out of that thread: the dictionary had a limit nobody was told about, so terms past a few dozen were being dropped silently. Those are now marked in Preferences. Shipping in 1.0.3. If you end up wanting this badly enough in daily use, tell me - a well-designed hotkey would change the maths and I'd revisit it.

@fatima__sani First of all thanks for commenting, this means a lot for a solo developer. And although we have a dictionary option, phantom voice doesn't learn from user corrections over time, and if by learning from user corrections you mean the app automatically checks what words are changed after the text are injected, then it would be a great feature for phantom voice to have, although at first glance it's gonna have some privacy issues since it's essentially monitoring your input field, and also we don't know how long it should monitor. If what you meant by learning corrections over time is something different, feel free to elaborate.

If you want lifetime access you can contact leftinverse@gmail.com for the phantom ID, i'll check that inbox from time to time. If you wanna try it out now, it's free to download on the site with 14 days trial no card needed: https://leftinverse.com/phantom/downloads

And if you have any other suggestions or other features you want phantom to add, feel free to reply. Yeah and also let me know if the indicator designs are any good, I spend a lot of time designing them and would like to know how people think about it.

@rabnoor_s You're right about the split, and it's the framing I've been missing.

On cleanup: it's already off by default in Phantom Voice, for close to your reason. It's a small local model, and a small model that meets a term it doesn't know will replace it with a fluent wrong one. The sentence still reads correctly, which is what makes it dangerous. So it's opt-in, and when it does run it's checked rather than trusted: it may remove fillers, fix punctuation and casing, or correct a mis-transcription into a term from your dictionary, and if the output differs from what you said in any other way it's discarded.

But per target app is the right control and I don't have it. Raw into Cursor, cleaned up into Mail, decided by where the text is going rather than by a mode you have to remember to switch. That's the part worth building.

The per-project word list I like even more, and a file in the repo is the right shape for it. You write the terms, it's version controlled, and anyone else on the project gets the same list. Phantom Voice already biases whisper on a word list, so the mechanism is there. A hand-written list is also small enough to fit whisper's initial prompt, which a scraped one would not be.

The part I don't have a clean answer to yet is how the app knows which project you're in. It has Accessibility permission, so it can see the frontmost window, but going from a window title to "you are in this repo" is guesswork. Might end up being that you register the folders once and it matches from there.

I'd also separate your examples. npx, pnpm, useEffect, MCP aren't project specific, they're vocabulary every developer shares, so that ships as a built-in list with no setup at all. Repo and branch names are the per-project half.

With that said I appreciate your feedback, and I highly encourage you to go try phantom voice out, it's free to download, 14 days of free trial with no card, and because you gave me some insights, I want to give you lifetime access to it. if you want the lifetime access phantom ID, just contact leftinverse@gmail.com about it, if you just wanna try it out, its free to download from the site. Thanks again