Roadmap and Vision

Roadmap & Vision

The vision

YazSes starts as offline hold-to-talk dictation and grows into a voice-first interaction layer for the desktop — still 100% on-device, privacy-first, and with every advanced feature off by default. The bet: you should be able to drive your computer by voice — dictate, edit, command, transcribe meetings — without any of your audio ever leaving the machine.

Two things make that credible rather than aspirational:

• It already works today. The stable release is a solid, cross-platform dictation daemon people use daily.
• The frontier is real and versioned. The v2 developer-preview line is a long series of research → decision-record → tested-core waves. Every capability lands as a pure, well-covered, off-by-default module before it's ever promoted.

Where things stand

Track	Status
Stable	A cross-platform offline dictation + voice-command daemon. This is what you install.
Core reliability (landing)	Mic-change auto-heal, a top-bar tray icon, and a "no text target" guard so words are never typed into the wrong place.
v2 developer preview	A large, growing catalogue of opt-in capabilities across dictation, editing, accessibility, and meetings — all off by default, none auto-published over the stable line.

(For the authoritative, always-current detail see ROADMAP.md and docs/v2-features.md.)

The v2 frontier, at a glance

The developer-preview line explores what a fully-offline voice layer can do. A flavour of the directions (all opt-in):

• Dictation & editing — prosody-aware auto-punctuation, spoken edit mode ("change X to Y"), confidence-ink homophone re-pick, spoken code and math→LaTeX.
• Accessibility — whisper-aware mode, semantic endpointing, screen-reader-paced injection, Braille output, dysfluency-friendly collapse for stuttered/dysarthric speech.
• Meetings & media — hands-free whole-meeting capture with speaker diarization and optional local-LLM minutes; offline file→text transcription with speaker attribution.
• Perception (experimental, opt-in) — glance-to-pane targeting via the webcam (frames processed in-RAM, never stored), a cocktail-filter voice gate, and code-switch language routing.

The throughline: breadth without compromising privacy or stability — heavy features are dependency-isolated behind optional extras and stay dormant until you enable them.

How we decide what to build next

Direction is set in the open, and community signal is a primary input:

1. 💡 Ideas discussions — the most-upvoted ideas rise to the top and drive the queue.
2. Fit with the north star — fast, private, on-device, plug-and-play. On-device is the tiebreaker.
3. Stability first — a very stable free application is the point; reliability work outranks new frontier features.

Good places to jump in

You don't need to touch the frontier to have real impact:

• Docs, install, and DX — the highest-leverage first PRs. See good first issues.
• Platform parity — help make macOS/Windows as smooth as Linux.
• Languages — multilingual model docs, non-English tuning, and code-switching.
• A frontier feature you care about — pick one from the roadmap, open an Idea, and let's scope it together.

→ Contributing — Start Here