Ubuntu to Offer Native Dictation: Canonical Launches Local Speech Project Myna

Ubuntu Is Finally Getting Its Own Native Dictation

On June 17, Canonical announced something new called Project Myna — native speech‑to‑text for the Ubuntu desktop. The first version will ship with Ubuntu 26.10 (codename Stonking Stingray) this October, and is positioned as “a core component of the desktop experience.”

The name comes from the myna bird, a species that mimics human speech. This naming is very Canonical‑style, continuing the tradition from Snap, Mir, and others.

Let’s get the key point out of the way first: this is not a voice assistant — not Cortana / Siri / Copilot‑type functionality. Canonical has been “excessively” restrained in the first version — no voice commands, no desktop control, no translation, no automatic language detection. Press a shortcut key, speak, and the text is inserted into the application where your cursor is. That’s it.

At a time when everyone is piling into “AI Agents” and “desktop operation agents,” Ubuntu has chosen the opposite route, focusing on solidifying the basic dictation functionality first. This product philosophy is worth noting.

Everything Runs Locally

Technically, Myna uses an AI speech recognition model, but all recognition happens locally. Once the model is downloaded, it works offline.

This continues Canonical’s broader AI roadmap. Back in April, Canonical’s VP of Engineering Jon Seager wrote about Ubuntu’s AI integration approach — dividing it into “implicit” and “explicit” categories:

• Implicit AI: Models run in the background, quietly enhancing existing features. Speech‑to‑text, text‑to‑speech, and accessibility tools are priorities.
• Explicit AI: AI‑native features for users who opt in, strictly opt‑in.

Myna is clearly an extension of the first category — making dictation an infrastructure‑level capability rather than dropping in a chatbox.

Canonical’s privacy promises are straightforward:

• Microphone is only accessed when the user activates dictation
• Audio data is processed in memory and discarded immediately after
• No uploads to external services
• Model runs locally, usable offline

Compare this to Windows 11 Recall’s PR disaster last year — Canonical’s “privacy‑first” narrative has credibility. When Fedora announced a similar plan this May, the Linux community’s debate over whether open‑source OSes should embed AI was still fresh in memory. Canonical clearly learned the lesson: developers care most about data control and the ability to turn features off, and Myna avoids pitfalls on both points.

Modular Architecture with Hidden Ambition

Myna’s engineering is more intricate than it appears. The architecture is split into four independent components:

• Speech recognition (ASR model inference)
• User interaction (shortcuts, UI feedback)
• Dictation management (session state, context)
• Text injection (writing to the currently focused application)

This separation has obvious benefits: in the future, swapping the underlying model (e.g., from Whisper to a lighter local ASR) or expanding text injection from Wayland to X11 and then to other desktop environments (KDE, Cosmic) won’t require touching other modules.

The target platform for the first version is Wayland on Ubuntu, with GNOME as the main validation environment. This is expected — since Ubuntu 25.04, Wayland has been the default session, and X11 is essentially legacy baggage on the desktop side. Text injection is trickier on Wayland than X11 (due to sandboxing), and Canonical clearly wanted to avoid being hampered by old protocols from the start.

Source code and architecture documentation are already on GitHub under GPLv3. This means any downstream distribution (including Pop!_OS, Linux Mint, Elementary — Ubuntu derivatives) can directly use it, and even Fedora or openSUSE could fork it.

Why “Dictation” and Not “Assistant”

Worth elaborating here. Canonical desktop team’s Jean Baptiste Lallement said they want community feedback “before too many design decisions are locked in,” especially from users who rely on dictation, accessibility users, and those who have tinkered with speech recognition on Linux.

Translated: they don’t want to make another Cortana nobody uses.

The voice assistant direction — from Amazon Alexa to Apple Siri, to all those attempts at “PC Siri” (Mycroft has already shut down) — has proven one thing over the past decade: in desktop environments, most users actually don’t talk to their computers. Not in the office, not at home when a keyboard is available. But dictation is another matter entirely.

Dictation solves the input speed problem. A skilled typist types 60–80 words per minute, while speaking can easily exceed 150 words per minute. For:

• Journalists, writers, technical authors
• People who develop RSI (repetitive strain injury) from prolonged typing
• Users with vision or motor impairments
• Non‑native English speakers who speak more fluently than they write

Dictation is essential. macOS has had Dictation for more than a decade, Windows also has it. On Linux, support has long relied on community projects like nerd‑dictation, whisper‑typer — fragmented, with varying quality and high setup barriers.

Myna fills this gap, which is the right direction.

Who It’s Competing With

Here’s a cross‑section of current local speech‑to‑text options on Linux:

| Solution | Model | Integration Level | Maintenance Status | |----------|-------|-------------------|--------------------| | nerd-dictation | VOSK | CLI / self‑config | Community maintained | | whisper.cpp + various wrappers | Whisper | Requires self‑assembly | Active | | Speech Note | Whisper/others | Flatpak app | Active | | Numen | VOSK | Voice commands focused | Active | | Myna | Not disclosed (likely Whisper family) | System‑level native | Canonical official |

Myna’s true differentiation is not in the model — likely Whisper or faster‑whisper under the hood — but in system‑level native integration. Community projects can’t achieve this: it requires GNOME cooperation, Wayland protocol support, a native Settings panel, default installation, default shortcut, and default discoverability.

macOS Dictation’s usefulness is not because the model is powerful, but because it works in every text field, triggered with a double press of Caps Lock. Myna’s goal is clearly the same.

Roadmap and Potential Pitfalls

Canonical’s stated rollout:

1. Ubuntu 26.10: First release, intentionally narrow features, polishing core experience
2. Subsequent versions: Continuous improvement of desktop integration, exploring more natural and accurate dictation
3. Priorities: Based on early user feedback

Some potential pitfalls:

Model size and first‑time experience: Local ASR models are often hundreds of MB to several GB; Whisper Large is nearly 3GB. Ubuntu installation media is already over 5GB — bundling a large model will inflate it further; downloading on first run may be criticized as disruptive. How Canonical balances this is worth watching.

Support for Chinese and other non‑English languages: There’s no mention of language coverage. Multilingual Whisper versions support Chinese fairly well, but smaller local models often have lower accuracy for Chinese than English. If the first version is optimized only for English (likely), Chinese Linux users won’t benefit much.

Conflicts with input methods: Linux desktop input method frameworks (IBus, Fcitx5) are longstanding headaches. How Myna’s text injection integrates with input methods is a technical detail prone to bugs.

Resource usage: Local inference isn’t friendly to low‑spec devices. Ubuntu prides itself on running on old hardware — if Myna spins up fans on an 8GB laptop, “core desktop experience” becomes a liability.

A Bigger Picture

Myna is just one part of Canonical’s AI roadmap. According to Jon Seager, Ubuntu 26.10 will also explore agentic workflows — such as automated troubleshooting (“an Agent fixing your system”). But like Myna, everything will be strictly opt‑in.

Canonical’s stance is clear: AI is about augmenting capabilities, not forcibly binding them. This sharply contrasts with Microsoft’s aggressive Windows 11 approach, and diverges from Apple Intelligence’s “bundle and sell” strategy.

For developers, this restraint is positive. Ubuntu, as the default OS for many developer workstations and servers, has no need for models secretly running in the background, sending data, or hogging resources. Canonical clearly understands this.

Whether Myna becomes a daily tool for Ubuntu users or another unused “built‑in app” will depend on its completeness at Ubuntu 26.10’s release. The GitHub repo is already open — those who want to give feedback can try it early.

Incidentally, speech app developers wanting to integrate stronger cloud‑based ASR or multi‑model comparison can use OpenAI Hub (openai-hub.com) — one key accesses APIs for GPT, Claude, Gemini, DeepSeek, and other mainstream models in an OpenAI‑compatible format, with direct domestic connectivity, saving account and networking hassles.

References

• IT Home: Canonical launches local speech‑to‑text project Myna — First‑hand Chinese coverage of Project Myna
• iThome Canonical tag page — Summary of Fedora and Ubuntu local generative AI integration plans and community reactions