Roadmap¶

YazSes is under active development. It is a fully-offline, hold-to-talk voice dictation daemon for Linux, macOS, and Windows — no cloud, no account, nothing leaves your machine. This page is the honest, user-facing picture of where the project stands: what has already shipped, what is being connected to the live dictation pipeline right now, and what is planned versus merely designed.

Three principles hold across everything below:

• Offline first. Transcription runs on your CPU with faster-whisper. There is no telemetry and no network dependency for dictation.
• Off by default. The stable dictation path is small and predictable. The large catalogue of extra capabilities is opt-in — you enable exactly what you want with yazses features enable <name> and apply with yazses restart.
• Nothing you didn't ask for. New capabilities land dormant and are activated incrementally, so an upgrade never changes how your dictation behaves unless you turn something on.

For the full catalogue of individual capabilities — what each does and how to enable it — see the Feature Reference.

Shipped¶

The stable release is v1.4.1. The table below groups milestones by minor version and gives the headline capability each one added; see CHANGELOG.md for per-release detail.

In progress / near-term¶

The current focus is wiring the large catalogue of built capability cores into the live dictation pipeline, one batch at a time.

Over a long series of research rounds, roughly 135 capabilities have been designed, implemented as self-contained cores, tested, and registered in yazses features. Many of these cores were built and verified but were not yet connected to any runtime path — so registering them was not the same as being able to use them. Closing that gap is the near-term work: each batch connects a few more cores to the real dictation flow, with a test proving each one fires only when you enable it. Everything stays off by default, so existing behaviour is unchanged.

Concretely, the batches wired in so far include:

• Live DICTATE-path text transforms — grammar repair, diacritic restoration, semantic line breaks, confusable-glyph warnings, inline compute, auto-pairing of brackets, phonetic correction of mis-heard names, transliteration, and structured-markup dictation now run in-pipeline when enabled.
• Offline text-tool commands — yazses reflow (monologue → outline), yazses table (spoken rows → CSV), yazses shellpipe (spoken pipeline → shell command text, never executed), and yazses braille (text → Unicode Braille).
• Offline recording import — yazses transcribe <file> decodes any common audio format, transcribes offline on CPU, and optionally tags who said what (--diarize), writing a .txt/.md/.srt/.vtt/.json sidecar. Speaker naming is opt-in and consent-gated; it never auto-enrolls other people.

Alongside this, recent near-term work has focused on making installation and first-run honest and unmissable: yazses setup now prints an ordered "finish installing" checklist, yazses start/restart verify the daemon actually came up (instead of reporting success while it crashed), and yazses doctor ends with a plain-language verdict and the exact next command to run.

The remaining cores will keep activating incrementally in the same way. A subset of the catalogue is live in the DICTATE pipeline today; the rest come online batch by batch.

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. An opt-in path to send audio to a hosted provider for a hard recording was fully designed, with strict guardrails — but it is not implemented, and offline remains the only path. It stays deferred so the "nothing leaves your machine" default is never quietly weakened.
• Personal speech adapters (LoRA). On-device fine-tuning that adapts the model to your voice and vocabulary — including atypical speech (dysarthria, ALS, Parkinson's) — is designed and gated on a measured accuracy win on held-out data, so it only ever applies when it genuinely helps. Prompt-level personalization from your own corpus already ships; the training-based adapter is the deferred part.
• Per-language model auto-switching. Detecting the spoken language and hot- swapping to the right model is designed as a routing layer; the language-specific model files are deferred.
• Code-switch dictation. Routing for mixed-language speech is in place, but stays dormant until a suitable code-switch model is supplied out of band.

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

• Silent-speech input (sEMG). Fully silent dictation via a surface-EMG wristband is designed; it activates when the hardware is present.
• Gaze- and vision-based targeting. Look-to-pane dictation routing and pure- vision screen commanding are designed, but depend on a webcam plus platform support that is not universally available today.

Speculative / research directions:

• Deeper multi-step reasoning and agentic task chaining over long sessions.
• Semantic retrieval (RAG) over your personal note corpus, fully offline.

These research directions are explored as fresh state-of-the-art rounds and may or may not become shipped features. When they do, they follow the same rule as everything else: off by default, opt-in, on-device.

Known limitations¶

A short, honest list of current constraints. For hands-on help with specific problems, see docs/troubleshooting.md.

• CPU transcription latency. Whisper runs on your CPU (int8). It is fully offline and private, but on a first run the model can take roughly 10–30 seconds to load, and larger/more-accurate models trade latency for quality. Choose a model to suit your machine.
• English-tuned by default. The default configuration targets English (small.en / base.en). Other languages work, but per-language models and code-switch support are still in progress (see Future work).
• Desktop-only. YazSes targets Linux, macOS, and Windows desktops. There is no mobile or web version.
• Some capabilities need optional extras or hardware. Heavier features (speaker diarization, gaze, EMG, neural denoise, and others) depend on optional Python extras or external hardware that are not bundled with the base install. They stay dormant until you install what they need 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 hotkey does not fire there — 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 on the project's Issues tracker.

The quickest way to find the exact link (and the current version you're running) is to run:

yazses about

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