p.01 · Dictation you can trust · Mac · Free beta
0downloads so far
live from GitHub · every release · how we count
Desi Dictation
Hold a key, speak the way you actually talk, release. English, Roman-script Hinglish or हिन्दी lands in whatever app you are typing in.
It checks with you only when it’s unsure, and never swaps a word on a guess. No cloud, no account: your voice never leaves your Mac.
Recommended · One line in Terminal
Paste one line. No warnings.
Open Terminal (⌘ Space, type Terminal), paste this and press Enter. It downloads the newest release, checks its SHA-256, puts it in Applications and opens it. Apple’s “could not verify” never shows up this way.
$ curl -fsSL https://raw.githubusercontent.com/s-samarth/desi-dictation/main/install.sh | bash
same app, same checksum, zero clicks.
Or the DMG
Heads up · macOS will cry wolf
Open the DMG copy and macOS says Apple “could not verify” it, and offers Move to Trash. Don’t. That warning is a bit bogus: macOS shows it for every app whose maker hasn’t paid Apple $99 a year for notarisation. This is a free, open-source side project, so I haven’t, yet.
Drag the app into Applications, then paste this into Terminal once:
$ xattr -dr com.apple.quarantine "/Applications/Desi Dictation.app"
It removes the “downloaded from the internet” flag from Desi Dictation and nothing else: no password, no system setting changed. Not sure? Paste it into Claude or any AI assistant and ask whether it’s safe.
$99 a year is a lot of chai.
p.02 · Watch · 75 seconds
The whole idea, in 75 seconds.
Most dictation apps guess, and you find out after you hit send. This one checks with you, only when it’s unsure. Sound on: it’s cut to the beat.
p.03 · Languages
Three languages, one key.
Dictation tools understand English or Hindi. Nobody in Bangalore talks like that. “Kal meeting hai, please deck ready rakhna” is one sentence, and it should land as one sentence.
Hinglish
You say
“kal meeting hai, please deck ready rakhna”
It types
Kal meeting hai, please deck ready rakhna.
Model · Oriserve Hinglish Prime
English
You say
“send the invoice by friday, and cc finance”
It types
Send the invoice by Friday, and cc finance.
Model · NVIDIA Parakeet · Qwen3-ASR
हिन्दी
You say
“kal meeting hai”
It types
कल मीटिंग है
Model · Vaani Hindi
Examples are illustrative. Every mode runs fully on your Mac.
p.04 · Features
Everything around the words.
Dictation is the core. The rest is what makes it yours, and what makes it safe to trust: it checks the words it’s unsure of, learns your names from the fixes you already make, and never swaps a word on a guess. All of it on your Mac.
01
Works in every app
WhatsApp Web, Slack, Mail, Notes, your IDE. If it has a text cursor, the words land there, and on your clipboard too.
02
Checks only when it’s unsure
Turn on Review before paste and a clean dictation still pastes at once. When a word is shaky, a panel shows the text with only those words underlined. Press 1 to fix, Enter to paste.
03
Never swaps a word on a guess
A name it isn’t sure of is offered, not written over. Only what you told it changes on its own: the words you saved, and um and uh.
04
Ums out, second thoughts caught
Um, uh and hmm just go. In Review, “meet at 2, no, make it 3” and “scratch that” show up as a suggestion you accept with one key; with the context checker, “like” and “you know” are offered as fillers too.
05
Learns from your fixes
It wrote Rowan, you fix it to Rohan in the text. A small popup asks “Always write Rowan as Rohan?” Nothing is learned without your yes.
06
Claude, but only when it fits
Some words are right only sometimes. Set one to When it fits and a small local model reads the sentence: “ask cloud to review it” becomes Claude, “upload to the cloud” stays cloud.
07
Your words, two lists
English and Hinglish keep separate word lists, so “nahi” never touches your English. Select any word in any app and right-click → Always write as… to add one.
08
Your Hinglish spelling
Hinglish has no fixed spelling. If the chat already says “nahi”, a dictated “nahin” is offered as “nahi”, so the message reads consistent.
09
Per-app language
Pin WhatsApp to Hinglish and Mail to English once. It switches by itself, and tells you when a rule kicked in.
10
Models from the menu bar
Switch between Parakeet (fastest) and Qwen (most accurate) without opening a window. Each language remembers its own model.
11
Pauses are fine
Voice-activity detection handles thinking pauses and long dictations. Transcription runs in chunks while you speak.
12
A day of history
The last 24 hours of dictations, searchable, stored only on your Mac. Closing the review panel without pasting still keeps it here. Turn it off if you like.
Not sure what to switch on? The app’s Guide explains every switch, right next to it.
p.05 · Review · Trust
It asks only when it’s unsure.
Dictation you can trust isn’t dictation that’s never wrong. It’s dictation that tells you where it might be, before it pastes, and never swaps a word on a guess.
When the panel opens
Dictation → Review → Review before paste
- Offdefault
Pastes straight away, as it always has. - Only when unsurerecommended
The panel opens only when a word is underlined. A clean dictation pastes with no extra step. - Always
Every dictation stops in the panel first. For when it really matters.
What the underlines mean
- Hard to hear (orange). The speech model wasn’t sure of the word, or it sounds like one of your words.Needs: Nothing extra.
- Odd here (purple). Heard clearly, but it doesn’t fit the sentence: “watch one PC”.Needs: The context checker. English and हिन्दी.
Nothing is changed for you. Hinglish uses orange only.
Keys
- Tab
- next underlined word
- 1–5
- pick a suggestion
- 0
- looks right: keeps the word, drops the underline
- Enter
- paste into the app you were in
- Esc
- copy instead of pasting
- ✕ or ⌘W
- close: no paste, clipboard untouched, still in History
How strict? Match it to the stakes.
Dictation → Review → Underline words. Stricter catches more mistakes but asks about more fine words. Start stricter, and loosen it as you learn where it slips.
Conservative
Underlines about 1 word in 7
Caught 56% of wrong words · 35% of underlines were real mistakes
Pick it for
- Email, documents, anything a person reads with your name on it.
- Names and numbers that must be right.
- Your first week, while you learn where it slips.
Moderate
DefaultUnderlines about 1 word in 12
Caught 43% of wrong words · 49% of underlines were real mistakes
Pick it for
- Everyday messages: Slack, WhatsApp, quick replies.
- When you’re not sure. It’s the default.
Liberal
Underlines about 1 word in 25
Caught 28% of wrong words · 67% of underlines were real mistakes
Pick it for
- Talking to a coding agent or a chatbot. It reads past a typo, so stop only for words that are probably wrong.
- Once you trust it and want speed.
Measured on English dictation with Parakeet on an M3 MacBook Air, context checker installed. Other models and languages land nearby; Qwen gives no confidence, so its underlines come from the context checker alone.
a review that cries wolf gets switched off, so Moderate is the default.
p.06 · Guide · Switches
Every switch, and when to flip it.
Everything past plain dictation is optional, and starts in its most careful state. Here is what each switch does, so you can turn on only what earns its place.
Optional download · 397 MB
The context checker
A small language model (Qwen3 0.6B) that runs on your Mac. It reads a sentence and judges which words fit. It never writes anything and never sends anything.
What it adds
- Purple “odd here” underlines, for words heard clearly but wrong in context.
- “When it fits” words: cloud becomes Claude only when the sentence is about Claude.
- Fillers judged: “like” and “you know” are offered for removal only when the sentence reads better without them.
- The best suggestion listed first, and which words “no, make it 3” replaces.
Everything else works without it: orange underlines, your words, um and uh removal, spoken corrections in review.
Where · cost
Models page → Context checker → Get. The same row sits under the Review settings and on the Vocabulary page. It uses about 0.4 GB while it works.
Worth it if you mostly dictate English with Review on, or your names sound like everyday words.
| Switch | Starts | Flip it when |
|---|---|---|
| Review before pasteDictation → Review | Off | Set Only when unsure once getting it right matters more than one keypress. |
| Underline wordsDictation → Review | Moderate | Conservative for email and documents, Liberal for chats and coding agents. |
| Remove um and uhDictation → Clean-up | On | Off if you want a word-for-word transcript. |
| Remove fillers when sureDictation → Clean-up | Off | On after a week of saying yes to its “Filler?” suggestions. Needs the context checker. |
| Suggest words from my correctionsVocabulary | Off | On if it keeps mishearing the same names. It asks before learning anything. |
| Suggest near-misses of my wordsVocabulary | Off | On once you have a few names in Your words: “Saraswath” gets offered as “Saraswat” in Review. |
| Suggest names on screenVocabulary | Off | On if you reply to people by name. It reads the window for that one dictation, then forgets it. |
| Remember language per appDictation → Advanced | On | Set a rule from the menu bar’s “For <app>” item: WhatsApp → Hinglish, Mail → English. |
| Keep historyHistory | On | Off if you’d rather nothing is kept, not even for a day. |
the app’s Guide page explains these too, next to the switches.
p.07 · Privacy · Speed
Your voice never leaves your Mac.
- No account, no sign-in, no email.
- No cloud transcription. Audio never leaves your Mac.
- No telemetry, not even the anonymous kind.
- It only goes online when you ask it to: downloading a model, or sending feedback.
- Names on screen, if you allow it, are read for one dictation and dropped. Never saved.
- The context checker is a model on your Mac too. It reads your sentence; it never sends it.
The models run on your Mac itself, so it is fast without being anyone else’s computer.
Measured, not promised
- 0.1 s
- per English dictation, on an M3 MacBook Air
- 1.8 s
- per Hinglish dictation, same Mac
- 0 bytes
- of your voice sent anywhere
p.08 · Compare
How it stacks up, honestly.
Wispr Flow, MacWhisper and Superwhisper are good apps. None of them is built for the way India actually talks, and that is the gap this fills.
| As of July 2026 | Desi Dictation | Wispr Flow | MacWhisper | Superwhisper |
|---|---|---|---|---|
| Hinglish in Roman script | Built for it | Not specialised | Not specialised | Not specialised |
| Where your voice goes | Stays on your Mac | Their cloud | Stays on your Mac | Stays on your Mac |
| Works with no internet | Yes | No | Yes | Yes |
| Learns your names | From your fixes, asks first | From your fixes, automatically | Replacements you set up | Words you add by hand |
| Shows the words it’s unsure of | Underlines them before pasting | Not in its feature list | Not in its feature list | Not in its feature list |
| Changes your words | Only what you set up; the rest is suggested | AI edits automatically; reviewers say it changes phrasing | Raw text unless you add an AI prompt | Whatever your mode’s AI prompt does |
| Filler words (um, uh) | Removed; “like”, “you know” offered | Removed by AI Auto Edits | Only with an AI prompt (API key or Ollama) | Through an AI mode |
| Free tier | Everything, during the beta | 2,000 words a week | Smaller models only | Limited, short history |
| Paid plans | None yet. No subscription. | $12–15 a month | €59 once | $8.49 a month, or $249.99 once |
| Languages | Hinglish, English, हिन्दी | 100+ | 100+ | 100+ |
| Runs on | Apple Silicon Macs | Mac, Windows, iPhone, Android | Mac (Intel too), iOS app | Mac, Windows, iOS |
From each product’s public pages and 2026 reviews, checked July 2026. Prices and plans change; check their sites.
where they’re better, said plainly:
- Wispr Flow’s automatic edits mean less to fix by hand, if you’re happy for it to rephrase you.
- Wispr Flow is on every platform, including Android, and its polish is the category benchmark.
- MacWhisper is years more mature and far better at transcribing files and meetings.
- Superwhisper has the deepest system of custom modes and runs on Windows.
p.09 · Pricing
Free while it’s in beta.
Free beta
₹0
- Every model: Hinglish, English, हिन्दी and the pause detector
- Every feature: Review before paste, the context checker, vocabulary that learns from your fixes
- No word limits, no account, no card
macOS 14+ · Apple Silicon
no subscriptions: the roadmap lists them as a non-goal.
p.10 · FAQ
Fair questions.
+Is it really offline?
Yes. Once the models are downloaded, dictation works with Wi-Fi off, vocabulary learning included.
+How does it learn my words?
Turn on “Suggest words from my corrections”, dictate, then fix a misheard word in the text within 2 minutes. When you stop typing (or press Enter), it asks whether to always write it your way. In Review you can also keep a fix Just this once, When it fits, or Always. It only asks about sound-alike spellings, never grammar or rewrites, and nothing is saved without your yes.
+Does it ever change my words on its own?
Only in ways you set up: the words you saved to Your words, and um and uh (a switch, on by default). Everything else it suspects, a misheard name, a filler, “no, make it 3”, is shown as a suggestion and waits for you.
+Which underline setting should I pick?
Conservative for email and documents, where a wrong name costs you. Moderate (the default) for everyday messages. Liberal when you talk to a coding agent or a chatbot, which reads past a typo anyway. Start stricter, and loosen it as you learn where it slips.
+Do I need the context checker?
No. It’s an optional 397 MB model that adds purple “odd here” underlines, When it fits words and smarter filler suggestions. Everything else works without it. Get it from the Models page if you mostly dictate English with Review on.
+Does it read my screen?
Only if you turn on “Suggest names on screen”. Then, when you start dictating, it reads the names visible in that window so a reply to Rohan isn’t typed as Rowan. The text stays in memory for that one dictation, then it’s gone. All of it is off until you switch it on.
+Where does the download count come from?
From GitHub, which counts every download of every release file. The number at the top adds them up across all releases and all three ways in: the Download button, the one-line installer and the GitHub releases page. It refreshes about every 10 minutes. It counts downloads, not people, and it includes my own test downloads. Nothing on this site or in the app tracks you to get it.
+How do I send feedback?
Menu bar → Send Feedback…. Type what happened; adding your app version or your last dictation is a tick box, off for the dictation, and you see the exact text before it goes. It’s sent to this site only when you press Send, then forwarded to my inbox and not stored. No audio, no IDs. Prefer email? The same window has an Email instead button.
+Why does macOS say it can’t verify the app?
The beta is signed with my own certificate but not notarised by Apple yet; that comes before the public launch. Install with the one-line Terminal command and you never see the warning. From the DMG, the install guide shows the one-time fix.
+How do I update?
Run the one-line installer again: it quits the running copy and installs the newest release. Your settings, history, dictionary and models are kept.
+Will it run on my Mac?
It needs an Apple Silicon Mac (M1 or newer) on macOS 14 or later. Intel Macs aren’t supported.
+How much space does it take?
The app itself is small. Models are separate: about 1.5 GB for Hinglish plus English, another 1 GB if you add शुद्ध हिन्दी, 2 GB more for the most accurate English model, and 397 MB for the optional context checker.
+Does it cost anything?
Not during the beta. Every model and feature is free, and there are no word limits.
+Is it open source?
The app code is on GitHub. The Hinglish models are Oriserve’s, released under Apache 2.0, and run on whisper.cpp; English runs NVIDIA’s Parakeet or Alibaba’s Qwen3-ASR.
Ready? Kal meeting hai.
Easiest is the one-line Terminal install: no Apple warning. Taking the DMG? The same spot has the one command that clears it.
Free · Open source · On GitHub
If Desi Dictation saved you some typing, star the repo.
It’s free, there’s no account, and the app sends me nothing, so a star is the only way I know someone out there is using it. Shower some love: it’s what keeps me building.
made by Samarth Saraswat. questions? raise a ticket.