QUILL Transcribe – Product Requirements

Version 1.0.0. Status: implemented; releases 1 November 2026 (first public release). In RELEASED_APPS from 2026-10-16, so every public QuillVille menu, tray menu and companion-install offer lists it, and Quill Cast’s and Quill Radio’s Edit > Dictate Here shows in public builds. Its own QuillVille menu is five rows (QUILL, Quill Radio, Quill Weather, Quill Converter, Quill Inkwell on Ctrl+Alt+Shift+F7 to F11); it is last in QUILLVILLE_APP_ORDER, so no other app’s row changes key. A release blocker remains: the screen-reader pass in Outlook, Word, Notepad and the browsers (docs/qa/screen-reader-checks-2026-10.md).

The full product requirements are in the repository at docs/Product Requirement Documents and Specifications/quilltranscribe.md. This page records what version 1.0 does with the decisions that document left open, and what waits for a later version.

Decisions taken for 1.0

  1. Name. QUILL Transcribe; product identifier QuillTranscribe; app key transcribe. Its icon is a text caret with words arriving at it.

  2. One owner. Version 1.0 does not have a full service that QUILL, QUILL Lite and Inkwell talk to. It has one owner at a time instead: QUILL Transcribe holds an operating-system lock in the family data folder while it runs, and Quill Inkwell’s Dictate Anywhere key is passed to it. Inkwell takes the same lock only for the length of its own session, so the two never listen at once. Requests between them carry one word (toggle, start, stop or show) and never any words that were spoken.

  3. Recordings. Transcribe a Recording is included. It is QUILL’s own command and pipeline, unchanged; QUILL Transcribe adds only a window to review the result and copy it, save it, or open it in QUILL.

  4. Internet models. At 1.0.0, only OpenAI, which already had QUILL’s key storage, consent and network review. It is listed only when a key is saved and asks before anything is sent. Since 2026-10-16 (dictation design record section 23, D-31 to D-37), Dictation Settings’ Speech Providers… adds twelve speech companies with the person’s own account (Azure, Google, Amazon, Deepgram, AssemblyAI, Speechmatics, ElevenLabs, Gladia, Rev.ai, Groq, Mistral, IBM Watson), each with its own consent question, keys in Windows Credential Manager, and none recommended. Choose Model (Ctrl+M) does not list them yet; Dictation Settings does. Local models stay the default.

  5. Where it works. No list of supported programs is claimed yet. The list comes from testing with NVDA, JAWS and Narrator in Outlook, Word, Notepad and the three browsers.

  6. One microphone (after 1.0.0, 2026-10-13). A second operating-system lock, separate from the owner lock, is held by whichever of QUILL, QUILL Lite, QUILL Transcribe and Inkwell has the microphone open, and only for as long as it is open. A second start is refused with the name of the program listening. It is separate because QUILL Transcribe holds the owner lock for as long as it runs, and the editors must still dictate while it waits in the tray. QUILL Lite’s window is claimed as QUILL’s is, so QUILL Transcribe does not type into it. QUILL’s Locked Dictation (Ctrl+F9) takes the same lock while it records.

  7. The editors’ keys (after 1.0.0). A dictation command QUILL Transcribe shares with the editors answers the editors’ chord (family rule 2), and the 1.0 key stays as a second key (rule 5). The editors’ dictation commands it does not have are listed, each with its reason, in quill/core/transcribe/parity.py (rule 11).

  8. The working tick (after 1.0.0). A soft tick once a second while a heard phrase takes more than a second to recognise, and one sentence after four seconds, once a session. On by default here, off in the editors. What you hear after each phrase caps it: words only gives the sentence and no tick, nothing gives neither.

  9. Recent Phrases (after 1.0.0). The editors’ window and code, on Shift+F11. Insert Again types into the program you go back to, through the same keystroke path and checks as dictation. Memory only for now.

  10. In other programs (after 1.0.0). “insert snippet”, “insert abbreviation” and “paste clip” work, because they only insert. “What can I say” is a spoken sentence while dictating, and Dictation Commands lists exactly what works in other programs, with the person’s own phrases.

  11. Never lose what was said (after 1.0.0, D-5). Words held while paused, and a phrase whose typing stopped part-way, are kept on disk, sealed with Windows DPAPI, until they are typed or discarded; Recent Phrases lists them after this run’s phrases, with Discard. A finished recording’s transcript can be typed into a field (Review Transcript’s Type into a Field). The text history is opt-in (Keep a dictation history: off, 1, 7 or 30 days), text only, clearable, and deleted when turned off.

  12. Snippet blanks by voice (after 1.0.0, D-8). “QUILL insert snippet” asks for each blank and choice, then types the filled snippet as one phrase. “read that” reads the whole of the last insertion (D-18).

  13. The wake phrase and holding the key (after 1.0.0, 2026-10-15). Both are the editors’ settings, off by default. The wake phrase listens only while another program is in front, and never while QUILL Transcribe, QUILL, QUILL Lite or a program on the never-type list is; nothing heard is kept until the phrase is heard, and the state line and tray say “Listening for your wake phrase”. Holding follows the start and stop key itself, by asking Windows whether that one key is still down while a press is followed, with no keyboard hook. Because a system-wide key needs Ctrl or Alt, what is said while it is held is typed when it comes up.

  14. Feedback where you work (after 1.0.0, 2026-10-15). The words heard so far and the unspoken state changes go to a braille display only. Dictation Status is the editors’ Alt+F9. Dictation Status, Resume and Discard Held Words each have an optional system-wide key, empty until chosen, picked and checked like the others.

  15. Correcting in other programs (after 1.0.0, 2026-10-16, D-10). The correcting commands work on the words this session typed, and only those, as backspace-and-retype; the other program is never read. Each is refused with one sentence when the focus moved, when Windows saw a key or the mouse after dictation’s own typing (when, never which key), or, for the commands that retype, in Word and Outlook until the insertion harness shows they are safe. “Correct that” with a one-answer engine asks the other built-in engine for a second opinion (D-19), in memory only.

  16. Correcting yourself, styles and polish (after 1.0.0, 2026-10-16). The editors’ settings and code (dictation design record section 28): Fix it when I correct myself (off) fixes “Thursday, no, Friday” before anything is typed; Dictation style offers Polished only when the review could reach an AI; Allow polishing my dictation with AI when I ask (off) makes “QUILL polish that” backspace the last phrase into Review Dictation and send one fixed request there, after the review’s consent, never by itself. Learning a correction asks by voice (“QUILL yes”), never in a window, because a window would take the focus from the field; it follows “correct that” and “choose”, which work here on the session’s own words since decision 13.

  17. Learning and profiles (after 1.0.0, 2026-10-16, D-25 and D-23). Practice Dictation is the one window of QUILL Transcribe’s own that dictation writes into: its field is an editor-style target written with one undo step a phrase, never keystrokes, and only after the same refusals as any start. Its recognition test scores with the benchmark’s own word-error count. Five lessons use the family’s Tutorials window. Seven named profiles (Standard, Email, Document, Markdown, Chat, Coding, Verbatim) overlay the dictation settings; switching is manual only (menu, tray, More Dictation Settings, or by voice), and Lock Profile stops voice switching. In another program, a formatting profile types nothing for a heading and says why, because the other program’s formatting cannot be applied by keystrokes QUILL Transcribe can trust.

  18. One settings store, and Dictate Here (after 1.0.0, 2026-10-16). “Share dictation settings with every QuillVille program” (dict.md D-13) replaces the one-way hand-off as the way settings move: family_dictation.json in %LOCALAPPDATA%\QuillVille\Dictation, written when Dictation Settings is saved anywhere and read here on the 800 ms clock, behind one check box that is on for a new install and off for one whose settings were changed. Keys, the tray, the working tick, strict mode, the read-back and every consent stay local. The word list moved to the same folder, copied with the old file kept. Quill Cast’s and Quill Radio’s Edit > Dictate Here (D-16) send a start request, or launch QUILL Transcribe with --tray --dictate-here. Dictation design record, section 30.

  19. Model advice from real speech (after 1.0.0, 2026-10-22, D-22 and B-1). Met. The recommendation, tiers, speed factors and memory peaks come from the real-speech machine run recorded in docs/qa/dictation-real-speech-benchmark-2026-10-22.md (dictation design record, section 36), shared with QUILL and QUILL Lite through one engine and catalogue. English is ranked by the LibriVox held-out audiobooks: Parakeet Unified 3.2 percent of words wrong at 0.21 per thread, so it is recommended on 8 GB with four cores; Parakeet TDT v3 (3.8, 0.11) on 8 GB with two cores. Spanish is ranked by the mean of FLEURS Spanish and the Spanish audiobooks, only from models measured in Spanish: Parakeet TDT v3 (4.4 and 7.3) on 8 GB, against the built-in Whisper’s 30.5 and 27.0, which stays the pick on 4 GB with Moonshine for English. The memory cap is 600 MB under 7 GB and 1.6 GB under 15 GB (it was 1 GB), and never more than is free less 500 MB. Tiers: Parakeet TDT v3, Parakeet Unified and Whisper base.en Lightweight; Nemotron Balanced. The advice’s note says the figures come from real recordings on one computer.

  20. Removed models (after 1.0.0, 2026-10-22, dictation design record section 37, which replaces section 32.6’s deprecate-for-one-release policy). Met. Distil-Whisper small.en, Whisper small.en, Distil-Whisper medium.en, Whisper medium.en and Distil-Whisper large-v3 were each beaten by Parakeet Unified on the held-out audiobooks (5.9, 4.3, 6.3, 3.6 and 4.5 percent against 3.2), faster on one thread and in no more memory, and are removed from the shared catalogue in the same release (quill/core/windows_dictation/removed_models.py). Whisper base.en and every multilingual Whisper size stay. A saved choice of one becomes Parakeet Unified when that is already downloaded, otherwise the built-in model (engines.coerce_engine); nothing is downloaded. The change is recorded in the family’s shared dictation folder and said once, the next time dictation starts in any QuillVille program. Downloaded files are not deleted: the folders (whisper-small.en-int8, whisper-distil-small.en-int8, whisper-medium.en-int8, whisper-distil-medium.en-int8, whisper-distil-large-v3-int8) stay under %LOCALAPPDATA%\QuillVille\Dictation\models (portable: data\dictation\models) until the person deletes them.

  21. Moonshine base and long phrases (after 1.0.0, 2026-10-22, PR #1683). Met. It returned nothing for any phrase over about 8.5 seconds; long phrases are now heard in pieces of 8 seconds or less, cut at the quietest moment, and joined. Measured after the fix: 5.9 percent on the held-out audiobooks.

  22. Help > What’s New (build 3, 2026-10-25). Met. The Help menu has What’s New on Ctrl+Shift+F1, the key every standalone QuillVille app gives it: QUILL Transcribe’s changelog (docs/CHANGELOG) opened as HTML in the browser through the family’s shared quill/ui/whats_new.py. A build with no changelog says so rather than doing nothing.

Still to check by hand: speed, memory, screen-reader lag and battery on a real 4 GB Windows 10 computer with NVDA or JAWS running (B-8), and Whisper medium and large-v3-turbo, which were too slow to re-run and keep their earlier figures.

Waiting for a later version