Launched this week
Phantom Voice
Local push-to-talk dictation for Mac
115 followers
Local push-to-talk dictation for Mac
115 followers
Push-to-talk dictation that inserts in place: hold fn, speak, release, and the text appears in the app your cursor is already in. Transcription runs locally with whisper.cpp + Metal, so your audio never leaves the Mac. An optional on-device cleanup pass strips filler words and fixes punctuation. Most dictation apps show you a recording dot or a generic soundwave. This one has hand-animated indicators. $10 once, no subscription. MIT source. macOS 14+, Apple silicon.
Phantom Voice
Local plus push-to-talk is the combination I'd want — always-listening is the thing that makes people uneasy, and a held key is an unambiguous consent signal in a way that a wake word isn't.
The number I'd want to know is latency from key release to text appearing. Local models trade speed for privacy, and dictation is unusual in that a one-second delay is fine but a three-second one breaks your thinking — you've already moved on. Where does it land on an average Mac, and does model size let you trade accuracy for that if you need to?
Phantom Voice
@ark_y_k This is how I personally use this app, and I'll get to the "3 seconds breaks your thinking part" real soon. When I have this app on and start talking to AI, I dictate everything that is roughly over 3 words, so for: "ok", "commit and push", "continue", I just type it out cause it's muscle memory, with that said, sometimes I also use dictation for under 3 words, and because my talking duration is super short, transcribing is almost instant. As for longer dictations for example 10-20 seconds, it's roughly 1.2s at 10 seconds and a bit over 2s at 20(on my Macbook Air M1 16G), since decode time scales with how long you spoke rather than being a fixed cost. So your 3-second wall is real butit sits further out than you'd think - I'd have to talk for 25-30 seconds straight to hit it. And for my experience, it transcribing for 3 seconds or maybe a little more doesn't mess up your train of thought, since you already spoke everything you want the AI to know and to do, and are waiting for the AI to do the work anyways.
Real numbers from my machine, an M1, with the model resident:
- Short sentence (~5s of speech): about half a second from key release to
text appearing.
- 10s of speech: about 1.2s. That's ~9x realtime.
- It scales with how long you spoke, not a fixed cost, so you'd only hit a
3-second wait after speaking 25-30 seconds straight. Normal dictation
sits in your "one second is fine" band.
- First dictation after launch is slower while the model loads. After that
it stays in memory.
As for the models: you can choose which model to use in preferences - Base, Small, Medium, or Large v3 Turbo. Turbo is the default because its distilled decoder is both faster and more accurate than Medium, so if you want more speed the move is to change to Base or Small.
I really appreciate the reply for this, and no, this comment is not generated, it's actually the developer of the app speaking. I want to offer you fulltime access for phantom voice cause you commented and a comment on a new launch of first product means more than you think. Reach out to leftinverse@gmail.com so I can send you the phantom ID, or if there's any other way to contact you, you can let me know.
@willis_liao1 The 3-word threshold is a useful detail, and it says something about the product that you have one at all — it means the cost of switching modes is low enough that you're making the call constantly rather than committing to one input method for the session.
That's the thing I'd want to know as a new user, actually. Not "is dictation faster than typing," but where my own threshold would land after a week. If it's 3 words for you, that's a much stronger claim about latency than any benchmark number.
Does it hold when the target isn't an AI chat? Talking to a model is forgiving of loose phrasing in a way that a commit message or a Slack reply to a colleague isn't — I'd guess the threshold moves up when the output has to be presentable.
Phantom Voice
@ark_y_k Honestly, it moves up. With AI chat I don't care if the phrasing is loose because the model doesn't. For a Slack reply or a commit message I'm looking at the actual words, so the question stops being "is this over three words" and becomes "is this sentence worth reading back and fixing."
@willis_liao1 That's the sharper version of it. The threshold isn't about length at all — it's about whether I'm going to reread the output. Anything I'll edit afterward, I'd rather have typed correctly the first time. Anything a model will absorb regardless, dictation wins immediately.
Which means the honest pitch isn't "faster than typing," it's "faster for the half of your writing that nobody reads closely." That's a smaller claim and a much more defensible one.
Thanks for thinking it through out loud rather than just answering — that's most of what I wanted to know. Good luck with it.
the opt-in rewrite does more than save 1.1 gb — a cleanup model can turn a technical token it doesn't know into a fluent wrong one. keeping punctuation separate from semantic rewriting makes the risk visible.
Phantom Voice
@qifengzheng You're right, and there's nothing in the app that stops it today. The only thing protecting against it is a prompt telling the model to keep every word exactly as it is - which is a request, not a guarantee. If it swaps a term anyway, nothing notices. And a wrong word that reads well is worse than an obviously broken one, because that's the one that gets shipped.
What your framing made me see is that the rule is checkable. Cleanup is only ever allowed to do three things: delete fillers, delete stutters, and fix punctuation and capitals. So I can compare the words before and after, allow deletions from a known filler list plus corrections to the user's own dictionary, and throw the whole result away if any other word changed - inserting the raw transcript instead. Same single model, but the rewriting you're worried about becomes impossible rather than just discouraged. Noted properly, not just in this thread. Thanks - this is the most useful comment I've had today. Will start implementing it in the app today.
Phantom Voice
Guys go download it on the site and try for free! I need feedback! Already found something that needs to be updated and that is the cleanuppass, now if toggled on, it uses a local LLM to clean up, but the downside is it requires another 1.1GB of RAM. A better way to clean up is what someone here pointed out: most of it is just string work. Fillers and stutters are a regex, and fixing my mis-transcribed technical terms is edit-distance matching against the
dictionary you already typed in - neither needs a model resident in RAM. The only part that genuinely wants an LLM is punctuation and casing, and whisper already does most of that. So the plan is to make the cheap path
the default and keep the 1.1 GB model as an opt-in for people who want the
heavier rewriting.
I've lost count of how times I've opened a diction app, waited for it to connect, and forgetten what I wanted to say.This push-to-talk workflow feels so much more natural. Great work!
Phantom Voice
@saira_bano4 thank you so much for the comment, you should download it on the site try it out, I promise you you won't be disappointed, you can also give me feedback of anything, including the price!
Phantom Voice
Guys let me know if the indicator designs are anygood, or are there any improvements I can make with the app
Phantom Voice
Yo guys forgot to mention the app is free to donwload now on the site with 14 days of free trial, no card and comes with the white wave and retro fox indicators. Guys go download it test it and let me know what you think! this is an app that I use myself in my daily work, when I first made the dication app, I just used it to keep working on the designs with codex and claude, so it would definitely benifit yall!