```
Canonical has released Myna, its speech-to-text dictation project for Ubuntu Desktop, as an installable technology preview — putting local, offline voice transcription directly into the hands of desktop Linux users rather than routing it through a cloud service.
The announcement, published on Ubuntu's official blog, marks Myna's move from an introduced concept to something users can actually install and run. The feature provides real-time visual feedback during speech recognition, showing words as they are recognised. The project takes its name from the myna bird, well known for its ability to mimic speech.
Privacy first, accuracy second
The newsworthy element here is architectural, not acoustic. Myna processes audio on the local machine, which means recorded speech never has to leave the user's device to be transcribed. For anyone whose dictation includes confidential material — legal drafts, financial commentary, medical notes, internal strategy — that distinction matters more than raw word-error rates.
That framing has particular resonance under Hong Kong's Personal Data (Privacy) Ordinance, where data minimisation and care over cross-border transfers of personal data are standing compliance concerns. Recording audio and sending it to a third-party inference service raises both questions at once. A locally executed model sidesteps them by design rather than by policy: there is no remote transcript to secure, sub-processor to assess, or retention schedule to negotiate. Organisations evaluating dictation tools for regulated workloads will find it a simpler story to take to a compliance conversation than any accuracy benchmark.
To be clear, this is a general observation about where the architecture is useful — Canonical's announcement makes no reference to Hong Kong requirements, and no local organisation has publicly responded to the release.
Engines under the hood
Myna supports three speech-recognition engines: Parakeet, Whisper, and FunASR. This matters for two reasons. First, it means users can switch between engines depending on their accuracy, language, and hardware trade-offs — rather than being locked into a single model. Second, Whisper's inclusion is relevant to the language question below, since Whisper's extensive language support is one of its best-known attributes. Canonical has not confirmed, however, how much of that language surface Myna will expose at launch.
How it compares to existing options
Ubuntu users who want offline dictation today have largely self-assembled: Whisper-based pipelines, often via whisper.cpp or Python bindings, remain the popular choice for accuracy, but they typically require manual model downloads, terminal wrangling, and — in the case of hosted Whisper endpoints — exactly the network dependency Myna is designed to avoid. Vosk offers genuinely lightweight offline recognition with a broader hardware footprint, but at a noticeable cost to transcription quality on noisy audio or non-standard accents.
Myna's pitch is integration rather than raw model performance: a first-party option that ships as an Ubuntu feature, wired into the desktop experience, without requiring users to build their own stack. Canonical has not published comparative accuracy figures against either alternative, and the technology-preview label should be read as an invitation to test and shape the feature rather than as a claim of production readiness.
Availability and hardware
Myna is now installable on Ubuntu 24.04 LTS, Ubuntu 26.04 LTS, and Ubuntu 26.10, on amd64/amd64v3 hardware running GNOME on Wayland. This is a notably broader footprint than a 26.10-only feature — LTS users on 24.04 and 26.04 can evaluate the preview without upgrading. Because inference runs locally, performance will track the host hardware, but the announcement does not detail minimum or recommended specifications. Users on older or low-memory machines should approach the preview as an experiment rather than a guaranteed experience.
The language question
Canonical's announcement does not specify which languages Myna will launch with. Given that Whisper — which supports dozens of languages — is one of Myna's supported engines, it seems reasonable to expect broad coverage eventually, but Canonical has not confirmed which language packs will ship with the technology preview or on what timeline. For Hong Kong-based deployments, that uncertainty is likely to be the deciding factor between trialling Myna now and waiting for a later drop. Cantonese support, from our reading of the announcement, remains unconfirmed; teams whose dictation workload is English- or Mandarin-centric can proceed with more confidence, while those dependent on Cantonese output should watch for a follow-up before committing to a workflow.
Canonical has positioned the release as an early milestone, and the open-source community will effectively be the testing ground for what comes next.
Canonical 已正式推出 Myna——其為 Ubuntu Desktop 開發的 speech-to-text 聽寫項目，以可安裝的技術預覽版（technology preview）形式發布，讓桌面 Linux 用戶可直接在本機離線進行語音轉錄，毋須經由雲端服務處理。
Ubuntu 官方網誌發表的公告指出，Myna 已由早期概念介紹階段，發展為用戶可實際安裝及運行的產品。此功能在語音辨識過程中提供即時視覺反饋，即時顯示已辨識的字詞。項目名稱取自八哥鳥（myna bird），該鳥以善於模仿人類說話而知名。
私隱優先，準確度次之
今次發布最具新聞價值的部分在於架構，而非聲學表現。Myna 在用戶本機上處理音訊，意味著錄音內容從不需要離開用戶的裝置便可完成轉錄。對於聽寫內容涉及機密資料的人士而言——法律文件草稿、財經評論、醫療紀錄、內部策略——這一點比原始的字詞錯誤率（word-error rate）更為重要。
此一框架在香港《個人資料（私隱）條例》（PDPO）下尤具意義：資料最小化（data minimisation）及跨境轉移個人資料的審慎處理，向來是合規範疇的核心關注。將音訊錄音並傳送至第三方 inference 服務，會同時觸發這兩個問題。本機執行的模型從設計上而非政策上回避了這些風險：沒有遠端轉錄稿需要保護，沒有 sub-processor 需要評估，亦沒有資料保留安排需要協商。為受監管工作負載評估聽寫工具的機構，會發現這是一個比任何準確度基準都更易帶入合規對話的論述。
須要澄清的是，以上只是關於該架構在何處適用的一般觀察——Canonical 的公告並未提及香港法規要求，本地機構亦暫無公開回應此次發布。
底層支援三種辨識引擎
Myna 支援三種語音辨識引擎：Parakeet、Whisper 及 FunASR。這一點有兩重意義。首先，用戶可按準確度、語言及硬件需要在不同引擎之間切換，而非被鎖定於單一模型。其次，Whisper 的納入與下文的語言問題息息相關，因為 Whisper 廣泛的語言支持向來是其最受稱道的特點之一。不過，Canonical 尚未確認 Myna 首發時會開放 Whisper 多少程度的語言支援。
與現有方案的比較
今日需要離線聽寫功能的 Ubuntu 用戶，大多依靠自行組裝方案：以 Whisper 為基礎的 pipeline，通常透過 whisper.cpp 或 Python 綁定實現，仍是兼顧準確度的熱門選擇，但一般需要手動下載模型、在終端機反覆操作——至於託管的 Whisper endpoint，則正好引入 Myna 設計上所要避免的網絡依賴問題。Vosk 提供真正輕量的離線辨識方案，硬件兼容範圍更廣，但在嘈雜音訊或非標準口音下，轉錄質素會有明顯下降。
Myna 的賣點在於整合而非純粹模型效能：作為 Ubuntu 第一方功能推出，與桌面體驗深度結合，用戶毋須自行搭建技術堆疊。Canonical 尚未公布與上述兩種方案的比較準確度數據，而「技術預覽」這一標籤應理解為邀請用戶測試及塑造該功能，而非聲稱已達生產就緒水平。
可用性與硬件要求
Myna 現已可在 Ubuntu 24.04 LTS、Ubuntu 26.04 LTS 及 Ubuntu 26.10 上安裝，支援 amd64/amd64v3 硬件並執行 GNOME on Wayland。此可用範圍明顯較只限 26.10 的功能更廣——使用 24.04 及 26.04 的 LTS 用戶毋須升級系統即可評估此預覽版。由於 inference 在本機運行，效能取決於主機硬件，但公告並未詳列最低或建議規格。使用較舊或低記憶體電腦的用戶，宜將此預覽版視為一項實驗，而非保證良好的體驗。
語言問題
Canonical 的公告並未指明 Myna 首發支援哪些語言。鑑於 Whisper——支援數十種語言——是 Myna 支援的引擎之一，長遠而言出現廣泛語言覆蓋似屬合理，但 Canonical 尚未確認哪些語言套件（language packs）會隨技術預覽版推出，亦未提供時間表。對於香港部署而言，此項不確定性很可能成為即時試用 Myna 或等待後續版本的決定性因素。就我們對公告的解讀而言，粵語支援仍未獲確認；聽寫工作以英語或普通話為主的團隊可以更有信心地先行試用，而依賴粵語輸出的團隊則應等待後續公告再決定是否投入工作流程。
Canonical 將此次發布定位為早期里程碑，而開放源碼社群將實質上成為接下來發展方向的測試場。