Framer AI AgentsDesign and publish professional sites with AI
Local Parrot
Your voice. Never shared.
10 followers
Your voice. Never shared.
10 followers
LocalParrot is a local, offline voice-to-text app. Everything happens on your machine. LocalParrot never sends your audio anywhere: no cloud, no account, no internet connection required to use it. Turn off your WiFi and try it, it works exactly the same. That's not a limitation, that's the point.
how well does it handle things like technical jargon or proper names since everything runs locally without the cloud models backing it up?
@nevzaty1ns Honest answer, and I tested exactly this before replying.
Out of the box it nailed: Kubernetes, OAuth, cache invalidation, hotfix, rebase, CI pipeline, async mutex, ONNX, deadlock, tachycardia, ibuprofen, amoxicillin.
What it missed: truly niche terms and unseen proper names. Redis came out "Reddy's", WASAPI came out "Waysap i", and my own surname came out phonetically.
That's not really a local-versus-cloud gap, by the way: speech models, wherever they run, meet brand-new words phonetically for the first time. The standard fix across the industry is a custom vocabulary.
LocalParrot's version of that is a personal dictionary: a plain JSON file on your disk mapping how a term must come out (product names, libraries, colleagues, brand casing). I added 6 entries for the misses above and re-ran the outputs: 6 of 6 fixed, including multi-word terms. Add a word once and it's yours forever, applied instantly and offline, in a file you own and can read.
There's also an optional on-device AI cleanup pass (a small local LLM) for punctuation and filler words.
The "turn off your WiFi and try it" line is such a confident way to frame privacy as the core feature instead of a footnote. Also love that there is no signup wall standing between me and actually using the thing.
@talha03nb Thank you, that means a lot. That line really is the whole philosophy in one sentence: we wanted privacy to be a property of the architecture rather than a promise. There is no server, no account system, no telemetry, so "your voice never leaves" isn't something we have to uphold. It's simply how the thing is built.
The no-signup part follows from the same fact: with no backend, there is nothing to sign up to. Your settings are a file, your personal dictionary is a file, your audio never leaves RAM.
how does it handle different accents and background noise, since it's all on-device?
@betlf4xj Fair question, so instead of hand-waving I benchmarked both before answering.
Accents: I ran real recordings from the GMU Speech Accent Archive (same English paragraph, speakers with different native languages) through the exact model that ships. Word error rates:
US native 0%,
German 2.9%,
Hindi 4.3%,
Russian 5.8%,
Spanish and Arabic 7.2%,
Mandarin 11.6%.
The one bad case was a very halting, heavily accented speaker at ~30%, mostly from one dropped sentence. Interesting detail: on accented speech, most "errors" were the model faithfully writing what was actually pronounced.
Tip: if you have a strong accent, pinning your language in Settings instead of auto-detect adds stability.
Background noise: same philosophy. I mixed noise into a dictation at decreasing signal-to-noise ratios: at 20 dB and even 10 dB SNR (a genuinely noisy room) the transcript stayed character-for-character identical to the clean take; it only started dropping words at 5 dB, which is noise at more than half the volume of your voice. Day to day, push-to-talk means the mic is only open while you hold the key, and a neural VAD trims non-speech before transcription.