Installing Myna Local Speech-to-Text on Ubuntu
Myna uses a hotkey to dictate into the focused application’s text field. You do not need to upgrade to Ubuntu 26.10 to try it: Canonical’s October 12, 2026 technology preview also supports 24.04 and 26.04 LTS.
Recognition runs locally, without an internet connection once a model is downloaded; audio is not uploaded or stored. Text inserted into an application follows that application’s own storage and network behavior. Official preview announcement , privacy boundaries .
Check Your Environment
Official coverage currently includes Ubuntu 24.04 LTS, 26.04 LTS and 26.10 on amd64 / amd64v3 with GNOME Wayland. Do not assume other desktops, X11 or ARM have the same support.
lsb_release -ds
dpkg --print-architecture
printf '%s\n' "$XDG_SESSION_TYPE"
printf '%s\n' "$XDG_CURRENT_DESKTOP"The session should be wayland and the desktop should be GNOME. A single command does not establish an architecture variant; also check your installation source and hardware.
Myna is a technology preview. Language coverage, accuracy, hardware requirements and application compatibility depend on the selected engine. Review important text before saving or sending it.
Install Myna Settings
Use the Canonical Desktop Team PPA specified in the announcement:
sudo add-apt-repository ppa:canonical-desktop-team/myna
sudo apt update
sudo apt install myna-configOpen Myna settings from Activities. Follow setup to install the client and models, select a recognition engine and configure the hotkey. Installation and model downloads need a network connection; a download still in progress is not a recognition failure.
Choose a Recognition Engine
For Chinese dictation, try FunASR / SenseVoice-Small first. Myna currently recommends this family for Chinese, Cantonese, Japanese and Korean users. Official model recommendation notes .
| Current model | Chinese support and selection |
|---|---|
| FunASR / SenseVoice-Small | Supports Mandarin and Cantonese; currently recommended by Myna for Chinese users |
| Whisper | Choose a multilingual checkpoint; checkpoints ending in .en are English-only |
| Parakeet TDT 0.6B v3 | Supports 25 European languages, excluding Chinese; unsuitable for Chinese dictation |
Coverage sources: Myna’s official test plan , NVIDIA model card . Myna’s FunASR package uses SenseVoice-Small; capabilities of Fun-ASR-Nano or SenseVoice-Large do not automatically apply.
SenseVoice-Small currently produces Simplified Chinese. Traditional Chinese users should check conversion requirements; the recommendation notes explicitly say that using Whisper with a language hint instead has not been evaluated for those users.
On supported NVIDIA GPUs, settings can offer an accelerated engine. Additional accelerators are planned. Use the options actually offered rather than replacing your graphics-driver stack just for the preview. Official support scope .
First Use
- Open a regular local text editor and focus an editable field.
- Use the configured hotkey to speak, observing feedback and transcription.
- Check punctuation, numbers, abbreviations and technical terms, then try a browser field.
- After confirming the model has downloaded, disconnect networking and repeat in the local editor to check offline recognition.
Myna currently focuses on dictation and does not control the desktop through voice commands. Application input fields may behave differently. Feature scope .
Check Language and Application Behavior
What Published Model Evaluations Establish
Table 6 of the SenseVoice paper reports the following Common Voice results. Values are character error rates (CER, percentages; lower is better) from the model authors, rather than this site’s measurements or overall Myna desktop accuracy.
| Test set | SenseVoice-Small | Whisper-small | Whisper-large-v3 |
|---|---|---|---|
| Common Voice zh-CN | 10.78% | 19.60% | 12.55% |
| Common Voice yue | 7.09% | 38.97% | 10.41% |
Source: FunAudioLLM paper, Table 6 . These results support trying SenseVoice-Small first for Chinese, without promising the same results for your accent, microphone, technical names or mixed-language text. CER includes insertions, deletions and substitutions; do not turn 100% - CER into a claim about a user’s recognition accuracy.
Check Your Own Device
Compare models with the same text, recording the model name, device, microphone and approximate delay. The following are checks you can perform, rather than measured results:
| Check | Suggested input | What to observe |
|---|---|---|
| Everyday language | A short work message | Missing words, homophones and punctuation |
| Mixed Chinese and English | Sentences containing Ubuntu, Git and project names | Correct product and project names |
| Numbers | Dates, version numbers and amounts | Decimal points, units and separators |
| Continuous dictation | One minute of speech | Delay, sentence boundaries and repeated text |
| Applications | Local editor and browser fields | Focus stability and successful text insertion |
Try drafts before using it for work, especially for recipients, commands, amounts and version numbers.
Troubleshooting Missing Text
Confirm that the microphone records in system sound settings, then check model downloads and the chosen language. Retest in a regular text editor to rule out application focus or hotkey conflicts. If failure is limited to one application, record its version and reproduction steps.
Client and model sandbox permissions can also matter; inspect settings and actual error messages rather than disabling system security. Report issues to the official Myna repository , including the Ubuntu version, session type, engine and reproduction steps. Avoid attaching private audio.
Related guides: AI image and speech tools, 26.10 or LTS.