Canonical released a technology preview of Myna, a dictation tool for Ubuntu. Recognition runs locally after downloading a model. The announcement lists GNOME on Wayland with Ubuntu 24.04 LTS, 26.04 LTS, and 26.10 as supported environments. Ubuntu official announcement
The announcement describes choices including Parakeet, Whisper, and FunASR. Local processing is useful, but it does not imply identical accuracy or hardware requirements across languages and engines.
On this page
Transcription and the destination application are different stages
Voice input captures microphone audio, converts it to text through a model, and inserts that text into an application. Local recognition means the conversion runs on the computer. Initial model downloads and subsequent storage or transmission of the resulting text are separate stages. A web application receiving the text can still send it to its own service.
Do not infer that every piece of work data remains local merely because transcription is local. Review any settings for retained recordings or temporary files, and examine the destination application’s storage policy. In a shared workspace, microphone activation also matters because other people’s speech can enter the recording. Include starting and stopping capture in the practical evaluation.
Technical language exposes different recognition errors
Product names, version strings, IP addresses, and English abbreviations mixed with Korean grammar create different difficulties from everyday sentences. Spacing and number formatting can fail together. Rather than relying on one successful short sentence, test recurring vocabulary from actual work. Sanitize any sensitive information before using it in the evaluation material.
Also observe how the tool decides that a sentence has ended. Ending during a thinking pause or waiting too long after speech can make longer writing awkward. Include correction time when comparing voice input with typing. Recognition accuracy and perceived speed depend on the microphone and ambient noise as well as the chosen model.
Start with three groups of sentences
Prepare everyday sentences, explanations containing product names and abbreviations, and sentences containing numbers or paths. Read the same material in a quiet space and in the normal workspace, then record incorrect words and required corrections. This is a proposed evaluation method, not a measured claim about the Korean accuracy of any particular engine.
Do not directly execute transcribed commands or configuration text where an incorrect number could have significant consequences. Review the values in an editor and make execution a separate decision. For an initial workplace trial, notes and explanatory prose provide a manageable way to discover input-flow and accessibility issues before extending use to more consequential tasks.
Include these in a desktop trial
- Verify that the GNOME and Wayland session is supported.
- Compare recognition errors and completion latency across microphones.
- Record accuracy separately for ordinary sentences and technical terminology.
A technology preview is an evaluation and feedback stage. Expand its use after verifying the accuracy and accessibility your work requires.
Announcement reference date: 2026-10-12 · Checked: October 24, 2026
.jpg){kind=link}