Skip to content

v2 Voice-First Interaction Layer (preview)

YazSes v2 grows from a dictation daemon into a broader voice-first interaction layer — still 100% on-device and privacy-first. Every v2 feature is opt-in and off by default, so your day-to-day dictation is unchanged until you turn one on. v1.4.x remains the stable release; the features below are a developer preview (v2.0.0-dev).

Manage all of them with the switchboard:

yazses features                 # list every capability + on/off + how to toggle
yazses features enable <name>   # turn one on (writes your config)
yazses restart                  # apply

Experimental features refuse enable unless you add --force.

Dictation upgrades (Wave A)

Feature Toggle What it does
Confidence Ink confidence Flags words the recognizer was unsure about (from Whisper's own probabilities) so you can re-pick them by voice instead of re-dictating. Local only.
Prosody pause→sentence [prosody] pause_sentence_ms Inserts a period when you pause for a sentence-length gap.
Spoken Edit Mode spoken-edit Edit the last dictation by voice — "change their to there", "delete the last sentence". Command-key gated.
Context-Primed Dictation context Primes the recognizer with terms from the active window/selection so domain words are transcribed right. Read transiently, never stored.

New capabilities (Wave B)

Feature Toggle What it does
Personal Adapter personalize Biases the recognizer toward your jargon and names, mined from your local corpus. Nothing leaves the machine.
Spoken Recall & Scratch recall Search your past dictations (yazses recall <query>) and capture spoken notes-to-self (yazses scratch). Corpus-local.
True Code-Switch polyglot Routes mixed-language speech to a code-switch model (you supply the adapter).
Voice-to-Tool (Spoken MCP) agent Run allowlisted tools by voice via MCP; state-changing tools ask first. Needs the agent extra + a local planner model.
Voice Pilot (AT-SPI) pilot Drive the desktop by voice via the accessibility tree — "click Save", "focus the terminal". Reads labels only, no screenshots.

Experimental (Wave C — --force to enable)

Feature Toggle What it does
Accessibility Continuum continuum Whisper/Low-Effort Mode lowers the mic gate so quiet or effortful speech is still captured (no shouting).
Modality Role Router modality Assigns each input its fastest role (gaze→point, EMG→command, voice→dictation). Needs EMG/gaze hardware.
Gaze-Routed Dictation [gaze] route_dictation Sends the next dictation to the window you look at, with a focus fallback and a confirm for destructive actions. Needs a webcam + calibration (X11).
Glasses↔Desktop Bridge bridge Dictate from a paired phone/glasses; the desktop does STT + typing. Local link only.

Wave D (v2.1) — new frontier features (all off by default)

A fresh 2026 state-of-the-art round. Manage them with yazses features enable/disable.

Feature Toggle What it does
Speech Translation translate Dictate in another language, type English — uses Whisper's built-in translate (X→English, no extra download); other targets via the seamless backend.
Tone-Aware Formatting affect Adds !/? from your vocal tone (beyond pause punctuation). Conservative by default; needs the affect extra for tone detection.
Predictive Completion predict A tiny on-device model suggests the rest of your sentence; accept by voice. Needs the predict extra + a model.
Noise Suppression denoise Cleans background noise/echo before transcription so dictation works in noisy rooms. Needs the denoise extra (DeepFilterNet).
Meeting Scribe scribe On-device "who said what" transcript — you are tagged You, others Speaker N. Needs the scribe extra (diarization).
Ask My Notes (voice RAG) rag Ask a question by voice, get an answer grounded in and citing your own local notes/docs. Needs the rag extra.
Codec Streaming codec Routes decoding to a low-latency streaming neural-codec engine (Kyutai/Mimi). Needs the codec extra; English/French-centric.
Voice Guard (experimental) voiceguard Types only when the live speaker matches your enrolled voiceprint and the audio isn't a recording/synthetic. --force; needs the voiceguard extra.

Two more Wave D directions are designed but await hardware: silent-speech (sEMG) dictation by mouthing words, and pure-vision screen commanding for surfaces with no accessibility tree. See design/adr/adr-v2-023/024 (internal).

Wave E (v2.1) — more frontier features (all off by default)

Feature Toggle What it does
Hallucination Guard hallucination Drops Whisper's fabricated ghost text on silence/noise (the phantom "Thank you.") before it's typed.
Voice Snippets snippets Say a trigger ("insert my signature") to type a stored template.
Phonetic Corrector phonetic Fixes mis-heard proper nouns by sound ("Cuber Netties" → "Kubernetes").
Multi-User Profiles voiceprint multi_profile Loads each enrolled speaker's own vocab/hotkey/cleanup from their voiceprint.
Hands-Free Auto-Stop autostop Tap once and speak; recording auto-stops when you finish.
Voice Mouse Grid mousegrid Drive the cursor and click by voice via a numbered grid, on any pixels.
Spoken Code Mode code Dictate code: spoken symbols → punctuation, word-groups → cased identifiers.
Spoken Math (LaTeX) math Dictate equations → LaTeX ("x squared plus y squared" → x^{2} + y^{2}).
Wake-Word Activation (exp) wakeword Start dictation hands-free on a keyword. Always-listening, local-only, --force.
Vocal-Strain Guard voicehealth Advises a break when your voice shows rising strain over a session (advisory only).

The zero-touch bundle — Wake-Word + Auto-Stop + Voice Mouse Grid — composes into a complete hands-free operating mode.

Wave F — self-improvement, adaptation & new modalities (v2.2)

All OFF by default. Pure logic ships now; heavy models load only when you enable the feature. Toggle with yazses features enable <name>.

Feature features name What it does
Speaking Coach coach Private analytics of your dictation: filler rate, words-per-minute, vocabulary diversity.
Smart-Paste smartpaste Adapts injected syntax to the target app (markdown bullets, code casing, URL autolinking).
Audio-Anchored Scrubbing scrub Keeps word timestamps so you can replay what you said or re-dictate one word.
Dictation Reflow reflow Say "structure this" to rewrite a ramble into bullets + action items.
Acoustic Context Profiles acoustic_profiles Detects your environment (quiet/café/car/meeting) and auto-tunes the mic gate + denoise.
Mood Ledger sentiment Private speech-sentiment journal; labels stay in the encrypted corpus.
Pronunciation Feedback pronunciation Per-phoneme good/fair/poor practice scoring for accent/L2 training.
Personal Read-Back Voice readback_clone Read the transcript back in a clone of your own voice (permissive OpenVoice V2).
Gesture Chords gesture Bind multi-input chords (held key + nod / key / sEMG squeeze) to actions.
Two-Way Live Interpreter interpret Face-to-face mode: each turn is detected and translated into the other language.

Wave G — normalization, privacy & new modalities (v2.3)

All OFF by default. Pure logic ships now; heavy models load only when you enable the feature.

Feature features name What it does
Entity ITN itn Spoken emails/versions → written form, no command words ("john dot doe at gmail dot com").
Redaction Ink redaction Masks spoken secrets (cards via Luhn, SSNs, keys) before they're typed.
Field-Aware Dictation fieldaware Shapes output by the focused field; refuses to type into password fields.
Corpus Voiceprint Scrub corpus_scrub Speaker-anonymizes stored learning audio (keeps content, not voice).
Compose-in-Target-Language compose Speak your strongest language, inject in a target language.
Grammar Repair gec Minimal-edit article/agreement fixes for non-native dictation.
Screen-Grounded Dictation screengrounded Primes STT with names visible on screen so they transcribe right.
Head-Pointer headpointer Move + click the cursor by head pose; fully hands-free with voice.
Silent Lip-Reading lipread Dictate with no audible voice; a webcam reads your lips.
Sign-Language Input sign Deaf/HoH signers sign to the webcam; ASL is recognized and typed.

Wave H — normalization, structured input & awareness (v2.4)

All OFF by default. Pure logic ships now; heavy models load only when you enable the feature.

Feature features name What it does
Emoji & Symbol by Voice symbols Speak symbols/emoji/arrows → Unicode ("right arrow" → →).
Voice Unit Conversion convert "twenty miles in kilometers" → "32.19 kilometers", offline.
Spoken Temporal Normalizer temporal "next Friday" → a concrete date, from the local clock.
Mid-Utterance Self-Repair self_repair "email Sarah no I mean Sara" → "email Sara" before typing.
Spoken Spreadsheet spreadsheet Grid navigation + cell addressing ("next row", "go to B7").
Clipboard-History by Voice cliphistory Recall recent copies by voice ("the second thing I copied").
Ambient Audio-Event Guard audioguard Auto-pause/alert on a doorbell, phone, alarm or your name.
On-Device Condense condense Insert a tightened summary of your own long dictation.
Slot-Filling Dictation slotfill One utterance fills a named-field template.
Few-Shot Command Spotter cmdspotter Enrolled low-latency micro-commands ("send", "stop").

Wave I — safety, patterns & power tools (v2.5, in progress)

All OFF by default. Pure logic ships now; heavy models load only when you enable the feature.

Feature features name What it does
Terminal Command Safety Gate cmdsafety Holds a destructive terminal command until you say "confirm".
Spoken Regex Builder spokenregex Build search patterns by voice ("four digits dash two digits" → \d{4}-\d{2}).
Structured-Markup Dictation markup Speak lists/tables → Markdown/org ("bullet list: apples; oranges").
Document Find-and-Replace findreplace Edit the whole document by voice ("replace every utilise with use").
Hard Contextual Biasing hotwords Biases recognition toward your vocabulary with a hotword trie so rare names win.
Voice Window Management windowctl Hands-free desktop layout ("move window left half", "workspace 3").
Citation-by-Voice cite "cite Vaswani 2017" → a formatted citation from your local .bib, offline.
Per-Language Auto Switching langroute Detects the language you speak and hot-swaps to its model + ITN.
Adaptive Latency Governor latency Keeps dictation responsive under CPU load, faster when idle (speculative decoding).
Diarized Conversation Capture diarize Attributed multi-speaker Markdown (Alice: …) with rename-by-voice.

Wave J — precision input, safety & recovery (v2.6, in progress)

All OFF by default. Pure logic ships now; heavy models load only when you enable the feature.

Feature features name What it does
Phonetic Spelling Mode spelling NATO words → exact characters for passwords/codes ("capital alpha bravo double lima" → Abll).
Voice Git Choreographer gitvoice Safe git by voice: destructive ops wait for a spoken confirm; the undo is always spoken.
Confidence-Gated Re-Ask reask Holds an uncertain word and asks you to pick/repeat it instead of guessing (flagship).
Verbatim / Autoformat Toggle verbatim "dictate verbatim" freezes all formatting mid-burst; "resume formatting" restores it.
Self-Learning Correction Dictionary corrdict Auto-fixes the ASR errors you keep correcting, learned from your edits.
Voice Fuzzy File Open fileopen "open the mortgage notes" → fuzzy-matches and opens a local file.
Voice Jump-to-Symbol jump "go to line 240" / "jump to function tokenize" → the editor moves there.
Spoken Shell Pipeline Builder shellpipe Speak stages → render a shell pipeline as text; nothing runs until "run it".
Recording Import recimport yazses transcribe <file> — transcribe existing recordings offline (wav/mp3/m4a/ogg/flac/opus/mp4) to a sidecar .txt/.md/.srt/.vtt/.json. --diarize tags speakers ("Speaker 1:"); --names/--rename or an enrolled voiceprint name them. Diarizer = sherpa-onnx (diarization extra).
Crowd-Proof Dictation (experimental) crowdproof Reconstructs your voice out of overlapping babble before STT (open-plan/café).

Wave K — keyboard, compute & session tools (v2.7, in progress)

All OFF by default. Pure logic ships now; heavy models load only when you enable the feature.

Feature features name What it does
Chorded Shortcut Synthesis chords Say any keyboard shortcut ("press control shift P", "hit F5") — no macro needed.
Inline Compute compute "what's 15% of 240" → types 36, on-device (never eval).
Voice Case Transform casetransform Recase the selection ("make this snake_case", "camelCase", "SHOUT").
Auto-Pairing & Wrap autopair Balances brackets/quotes in dictated code; "wrap this in parens".
Voice Undo/Redo Timeline timeline Undo/redo YazSes output across bursts by voice, even where Ctrl+Z is unreliable.
Session Bookmarks bookmarks "bookmark here as intro" + "jump to my last bookmark" for long sessions.
Spoken Table Entry tablecsv "row: Ada, 1815, London" → tab/comma cells, "next row" advances.
Word-Count & Goal Tracker wordgoal Count dictated words, set a goal, ask "how many words so far?"
Local Voice Timer voicetimer "set a timer for 25 minutes" — offline, spoken by read-back.
Focus-Class Auto-Profile focusprofile Grammar profile follows the focused window's class (terminal→shell, editor→code).

Wave K is complete — all ten features (see design/vision/v2-research/13-wave-k.md) are shipped, pure, 100%-covered, and off by default.

Non-speech & prosodic interaction (Wave L)

A new area: non-speech vocal signals and acoustic prosody as interaction channels — for users who can phonate but not articulate discrete words, and for word-free eyes-free control.

Feature features name What it does
Vocal Joystick vocaljoystick Continuous analog cursor control by sustaining vowels ("eee"→up, louder=faster, pitch-jump=click). Experimental.
Earcon Feedback earcon Non-speech state tones (rising motif = recording, buzz = low confidence) — eyes-free, faster than read-back.
Beam-Steered Spatial VAD spatialvad With 2 mics, gates out sound not from your seat by its arrival direction (pure-numpy GCC-PHAT).
Prosodic Auto-Punctuation prosodypunct Inserts periods/commas/question marks from how you speak — no need to say "comma".
Hesitation-Hold Endpointing hesitation Holds the turn open on filled pauses ("uhh…") instead of cutting you off.
Pitch-Contour Gestures contour Hum a shape — rise=confirm, fall=cancel, rise-fall=undo. Word-free. Experimental.
Breath-Paced Dictation breath Segments dictation by your natural breath groups, not silence.
Whisper-Aware Mode whispermode Detects a whisper and adapts gain/VAD/prompt so quiet dictation stays accurate.
Mouth-Sound Switch Access mouthswitch Non-verbal mouth sounds drive a scan-and-select selector (AAC switch access). Experimental.
Involuntary-Vocalization Excision involuntary Drops cough/throat-clear/sneeze before they corrupt the transcript.

Wave L is complete — all ten features (see design/vision/v2-research/14-wave-l.md) are shipped, pure, 100%-covered, and off by default.

Minimal-bandwidth AAC & text-intelligence (Wave M)

The lowest-bandwidth input methods and new text-intelligence layers.

Feature features name What it does
Vocal Morse morsevox Type by Morse using two vocal sounds (short/long); adaptive timing. Experimental.
Checksum-Validated Entry checkdigit Verifies dictated account/ID numbers (Luhn/ISBN/Verhoeff) and suggests fixes.
Semantic Line Breaks sembr Breaks prose one clause per source line so git diffs stay clean.
Acronym & Glossary Manager acronyms Expands acronyms on first use, contracts after, warns on undefined.
Style-Consistency Enforcer styleguard Applies your house style sheet to each dictation (Vale-lite).
Suggestion-Mode Dictation suggestmode Emits edits as CriticMarkup tracked changes for later review.
Screenplay Auto-Format screenplay Formats dictated scenes/dialogue as Fountain screenplay markup.
Spaced-Repetition Capture srscap "remember that X is Y" → a local Anki cloze card (SM-2).
Diagrams-as-Code by Voice diagramvox Dictate a flowchart → Mermaid/Graphviz source, no canvas.
Interruptible Proofreading proofback Interrupt the read-back and land the cursor on the exact word.

Wave M is complete — all ten features (see design/vision/v2-research/15-wave-m.md) are shipped, pure, 100%-covered, and off by default.

Structural editing, i18n & accessibility-output (Wave N)

Structural code editing, internationalization, and accessibility-output correctness.

Feature features name What it does
HatSelect hatselect Address tokens by spoken label ("delete bravo") for structural editing. Experimental.
Transliteration translit Dictate a native language in Latin letters → native script (Finglish→Persian).
BrailleOut brailleout Emit dictation as Grade-2 UEB Unicode Braille for Braille displays.
Spoken Outline outline "New item / indent / promote / collapse" drives an outline tree → Markdown/OPML.
Diacritize diacritize Restore dropped diacritics — "cafe" → "café", "naive" → "naïve".
SafeGlyph safeglyph Flag Unicode confusables/invisibles in identifiers and URLs before injection.
WordFind wordfind Offline reverse dictionary — describe a word, get a ranked shortlist.
LoadGuard loadguard Widen confirmations / defer risky actions when speech shows rising cognitive load.
Echo echo "Play that back" replays your own captured audio for a text span (not TTS).
Screen-Reader Pacing srpace Pace injection to a screen reader's reading rate, clause-chunked.

Wave N is complete — all ten features (see design/vision/v2-research/16-wave-n.md) are shipped, pure, 100%-covered, and off by default.

Privacy

Every v2 feature honours the same guarantees as the rest of YazSes: on-device processing, no telemetry, and no transcript persistence unless you explicitly enable the encrypted learning corpus. Context and gaze signals are read transiently and never stored; the bridge stays on your local link. See the privacy statement.