Long Mac dictations now keep their last words

What went wrong

The problem showed up in longer recordings, around 35 to 45 seconds in the cases we traced, where the text came back cut off before your final sentence. There were two independent causes.

1. Voice Filter trimmed your closing words. Voice Filter removes other people’s voices from a recording. It splits audio into windows and checks each one against your enrolled voice. The last window of a recording is mostly silence padding, so it gives a weaker voice match. Sometimes the filter decided your own last few words belonged to “someone else” and removed them.
2. FluidAudio 0.15.6 dropped some endings. With the Parakeet speech model, the FluidAudio library could decode a whole stretch of audio to nothing. In some long recordings, that stretch was the end.

What 0.2.38 changes

Voice Filter now matches you by speaker identity across the whole recording, rather than judging each window on its own. It removes a voice only with real evidence that it is another person: at least three seconds of speech that was not first heard in the padded final window. A short closing phrase in your own voice stays in the text.

Plainsay now uses FluidAudio 0.15.8, which recovers the audio windows that previously came back empty. We added a test that runs 35 to 70 seconds of speech through the Parakeet engine and checks the words spoken after the 25-second mark. It fails on the old version and passes on the new one.

Nothing changes in how you use Plainsay. Update, keep dictating, and your endings stay where you said them.

Get Plainsay 0.2.38

Plainsay runs on Apple silicon Macs with macOS 14 or later. Local mode is free: no account, no subscription, and no word or minute quota. Speech recognition runs on your Mac after a one-time model download.

• Download the notarized DMGThe latest release, 0.2.38.
• Read the source on GitHubThe Mac app is MIT-licensed.

Or install with Homebrew:

brew install --cask conrader/plainsay/plainsay

Already installed? Plainsay checks for signed updates automatically, or choose Check for Updates… from its menu-bar menu. With Homebrew, run brew upgrade --cask plainsay.

Found an ending that still goes missing?

Plainsay is open source, so you can read exactly what changed and tell us if a recording still loses its tail. Open an issue on GitHub with the recording length, the speech model you use, and whether Voice Filter is on. That is the detail that made these fixes possible.

This page describes fixes in the released app. It is not an accuracy or speed benchmark.