Roadmap¶

YazSes is a fully-offline, hold-to-talk voice dictation daemon for Linux, macOS, and Windows — no cloud, no account, nothing leaves your machine. The current stable release is v2.15.1, published on PyPI, GitHub Releases, Snap, and the APT repository.

This page is two things at once: an honest status report, and a statement of where we are going. If you only read one thing, read the pictures.

The destination¶

The keyboard is a 150-year-old constraint, not a law of nature. YazSes exists to make it optional — to become the private layer between you and your computer that hears you, understands you, and acts for you, entirely on your own hardware. Not a cloud assistant that rents you access to your own words: an instrument you own, that gets better the more it learns about you, and that never tells anyone what it learned.

The finished shape has three faculties, wrapped in two promises:

mindmap
  root((YazSes))
    Perceive
      Voice through Whisper and Parakeet
      Gaze knows where you look
      Muscle signals for silent input
      Meetings and recordings
    Understand
      Your personal vocabulary
      Dictation or command intent
      Who said what
      Your languages mixed freely
    Act
      Type into any app
      Command the desktop
      Write the meeting minutes
      Recall your own words
    Trust
      Offline always
      Encrypted on device
      Off by default
    Everywhere
      Linux macOS Windows
      Android next
      Every package manager

The eras¶

Each era is a promise kept before the next one starts. Foundation is shipped. Perception is shipping now. Understanding is the active frontier. Recall is the reason the project exists.

timeline
    title From dictation tool to keyboard-optional computing
    section Foundation — shipped
        v0.x to v1.4 : Hold-to-talk dictation that just works : Linux, macOS and Windows : Encrypted on-device learning loop
    section Perception — shipping now
        v2.x : Meeting Mode with speaker diarization : Second STT engine (Parakeet TDT) : Gaze deixis and EMG activation : Sotto-voce command channel
    section Understanding — next
        v2.15+ : Settings GUI (no terminal needed) : Offline command mode over selected text : Vocabulary biasing in every engine : Android dictation keyboard
    section Recall — the horizon
        v3 : Spoken recall over everything you ever dictated : Ambient scribe you control : A computer you talk with, that stays yours

flowchart LR
    F["🏗️ Foundation<br><em>a tool that works</em>"] --> P["👁️ Perception<br><em>it hears and sees well</em>"]
    P --> U["🧠 Understanding<br><em>it knows what you mean</em>"]
    U --> R["💬 Recall<br><em>it remembers for you</em>"]
    style F stroke-width:1px
    style P stroke-width:3px
    style U stroke-dasharray: 5 5
    style R stroke-dasharray: 2 4

(solid = shipped · bold = in progress · dashed = ahead)

Working milestones — where you can join¶

Every open issue belongs to exactly one milestone, and every milestone is an outcome, not a bucket. Pick the one that sounds like you:

Start at the pinned contributor guide, or — for research-shaped work — at Students, researchers & industry.

Three principles that never move¶

• Offline first. Transcription runs on your CPU. No telemetry, no network dependency for dictation, ever. The one designed cloud-escalation path stays deliberately unbuilt so this promise is never quietly weakened.
• Off by default. The stable dictation path is small and predictable. The large catalogue of capabilities is opt-in: yazses features enable <name>, then yazses restart. An upgrade never changes behaviour you didn't ask for.
• Honest about what exists. yazses features distinguishes what is wired and working from what is designed-but-planned, and refuses to pretend otherwise. This roadmap follows the same rule.

Shipped¶

Future work¶

The items below are planned directions. We distinguish clearly between what is designed but deliberately not built yet and what is speculative.

Designed, but explicitly deferred:

• Cloud escalation for transcription. Fully designed with strict guardrails — and not implemented, so the "nothing leaves your machine" default is never quietly weakened. Offline remains the only path.
• Personal speech adapters (LoRA). On-device fine-tuning that adapts the model to your voice — including atypical speech (dysarthria, ALS, Parkinson's) — gated on a measured accuracy win on held-out data. Prompt-level personalization from your own corpus already ships.
• Per-language model auto-switching and code-switch dictation. The routing layers exist; the language-specific and code-switch models are the deferred part.

Hardware- or model-gated (designed, waiting on the missing piece):

• Silent-speech input (sEMG). The activation seam ships today; it comes alive when the wristband hardware is present.
• Vision-based screen commanding. Gaze targeting ships on X11; deeper pure-vision commanding depends on platform support that is not universal yet.

Speculative / research directions:

• Spoken recall — semantic retrieval (RAG) over your personal dictation corpus, fully offline. This is the Recall era's core bet.
• Deeper multi-step task chaining over long sessions.

When research directions ship, they follow the same rule as everything else: off by default, opt-in, on-device.

Known limitations¶

• CPU transcription latency. Whisper runs on your CPU (int8). First model load takes roughly 10–30 seconds, and larger models trade latency for accuracy. Choose a model to suit your machine.
• English-tuned by default. The default configuration targets English (small.en / base.en). Other languages work; per-language auto-switching and code-switch support are still in progress.
• Desktop-only, today. An Android app is in design — see the mobile programme for the architecture and how to help. iOS follows Android for a platform reason explained there.
• Some capabilities need optional extras or hardware. Diarization, gaze, EMG, neural denoise and others stay dormant until you install their extras and enable them.
• Linux packaging caveat. Install via the APT script or pipx for hold-to-talk dictation. The strictly-confined snap cannot read the keyboard device, so the snap only serves the offline file-transcription use case.

Requesting features and reporting bugs¶

Feedback drives what gets built next. To report a bug or request a feature, open a GitHub issue — or run:

yazses about

It prints the author, version, project links, and where to report issues. yazses doctor also ends with a contact footer, so if you hit a problem you always know where to go.