Parrot writes the transcript in whole sentences, not word soup. While someone is talking you'll see a light preview filling in, and the moment they pause, the finished sentence lands as a bubble. If someone talks for a very long stretch without pausing, Parrot commits what it has every 12 seconds so the transcript never falls far behind.
One honest note: a sentence can only be finished once the speaker actually pauses, so the final text always arrives a beat after the words. That beat is the price of getting clean, complete sentences instead of fragments. The preview keeps you company in the meantime.
Out of the box Parrot auto-detects the language per sentence, which is handy for mixed-language calls but can occasionally guess wrong on a noisy bit, and the preview can flicker through a wrong language before the real sentence lands. If your calls are always in one language, set it explicitly in Settings → Transcription and everything gets steadier.
Whisper has never met your company name. If the transcript keeps mangling names or jargon ("LaunchEase" becoming "Lawn Cheese" is a real example from this very app's history), add those words in Settings → Transcription → Custom vocabulary, separated by commas. Parrot whispers them to the model on every pass so it spells them right.
On-device Whisper is the default and needs no account. If you prefer, you can transcribe through Groq or stream through Deepgram with your own API keys in Settings → Transcription. If a cloud engine hiccups mid-call, Parrot quietly falls back to the on-device model for the affected part, so a flaky connection never costs you transcript.