A note from the maker
Today, developing an app isn't so much about the technology — it's about the
thought process behind it, and the time spent refining and perfecting what the product
puts out. That effort is what makes Fabled Flow what it is in privacy, speed, and
accuracy.
I created Fabled Flow because I wanted a non-cloud alternative: private, local, and
genuinely affordable. I tested many dictation products, and none met my speed and
accuracy requirements. So I built one that would — and benchmarked it until dictating
felt almost magical: your words appear in the window quickly and accurately, with no
audio ever sent to the cloud, and formatting only when you ask for it.
My goal is to make products that close a real gap while staying private and
accessible to as many people as possible, on as modest a computer as possible — which
is why Fabled Flow is proudly developed on a first-generation Citrus MacBook Neo. I
believe in local tools like Apple Intelligence, and where they fall short, in local
backup models that do the same job — sometimes a touch slower, sometimes covering
languages Apple's model doesn't. Everything below is the record of that refinement.
Since the first version on July 6th, Fabled Flow has been through
148 recorded rounds of refinement across more than 130 numbered releases — about
164 hours of recorded, hands-on build sessions in the first three and a half weeks
alone (the session transcripts are kept, so that number is measured, not guessed),
and the daily refinement has continued since.
Very little of that was adding features. Most of it was process — human decisions
about how dictation should behave, made by using the app hard and refusing to accept
"close enough." Some of what came out of it you won't find elsewhere: the end of
every long dictation gets a second listen, because trailing words are the easiest
words to lose — and when the engine can't hear the same last words twice, the app
tells you instead of pasting a guess. The microphone is genuinely off between
dictations — we took on real engineering pain to keep it that way rather than soften
the promise. Every dictation keeps its receipts, so when something goes wrong we can
prove where — engine, pipeline, or app — in seconds. Every refinement started life as
a real dictation that went wrong on a real day, kept verbatim as a test the product
must pass forever. And every rule the product applies to your words fits on one page,
on purpose — rule sprawl is the quiet killer of dictation tools. None of those calls
came from AI; they came from caring about what lands at your cursor. The record below
is what that looks like, release by release.
Privacy has a price, and I chose to pay it. Because your
voice never leaves your Mac, there's no cloud model to lean on — every gain in
accuracy had to be earned by hand, rule by rule, through decisions that have nothing to
do with coding and everything to do with how dictation should behave. When the app
smooths a stumble ("c… customer"), no language model did that — none can,
and there's no standard to download; it came from hours of testing real speech. Even the
features still to come are made this way: your Mac quietly notices — locally, only
for you — which sound-alike words you personally fix ("are", "our"), so that one
day the app can offer you corrections built from your own patterns instead of imposing
someone else's. And through all of it, every word you speak is kept: no rewriting, no
quiet deletions, a complete record. Try dictating
A47_MAN/59-138(B) in
any other dictation app — the difference you'll see is where all those hours
went.
10.4
4 September 2026 · build b2962
Mac · the tuned-binary batch, notarized
- A half-spoken last word is never guessed into a full one. When the recording cannot prove how the word ended, the fragment is dropped with the stumble cleaner on and kept exactly as heard with it off — never the engine’s guess.
- The paste card fires less, and only when a paste truly could not be verified. It no longer trusts an app’s wrapper element as proof that text landed, and remembers per app when its reading is blind.
- Phantom words at the end of a dictation are refused before they reach the text. The engine’s own weak guesses tied on one instant over silence are dropped — measured at 6 of 6 catches and 0 of 320 false alarms.
- A long recording that produced nothing is no longer a dead end. It lands in History held, with its audio kept and a “Transcribe as…” action that re-transcribes it in the language you choose.
- Diagnostics keep their evidence. Three generations of the debug log, with per-clip analysis lines in their own file, so an incident can no longer be rolled away within minutes.
- Notarized by Apple. macOS opens this build without any warning.
10.4
3 September 2026 · build b2899
Mac · the trailing-word tune, notarized
- The last word of a dictation can no longer vanish. When a dictation ends on a softly spoken word, the clip-tail guard used to take one short re-hearing of that tail and, if it came back empty, treated the silence as proof and dropped the word (“…continue the work” lost “work”; “…this error or not?” lost “not”). It now takes two hearings at different offsets and keeps any word either hearing backs; an all-empty result is no opinion and the word stands.
- Verified on every reported specimen. 7 of 7 lost words present in the replay, and on the full benchmark rig the empty-output rate on the TED-LIUM set fell from 40% to 2.0% raw / 0.7% edited with no regression on any other set.
- Notarized by Apple. macOS opens this build without any warning.
10.4
3 September 2026 · build b2883
Mac · the dictionary digit guard + the capture-safety pair, notarized
- Numbers you dictate are literal, always. A spoken version number like “5.1” can never again be rewritten into a dictionary word — number-bearing words are untouchable by the fuzzy dictionary unless your taught replacement keeps the very same digits. All your taught word corrections keep working exactly as before.
- A recording can no longer be interrupted by a report card. Cards now wait until the capture is safely finished, and a release-time net catches any capture left orphaned — no dictation is lost to a card again.
- Repeat cards stop stacking. The card loop is scoped to the actionable face, so the same issue no longer piles duplicate cards.
- Your error reports carry your own words. The History card and the admin view now show the description you filed, not a generic label.
- Quiet housekeeping. Clipboard restore hardened, diagnostics logs rotate daily so they never crowd the disk, and secure password fields are never captured, pasted into, or recorded.
- Notarized by Apple. macOS opens this build without any warning.
10.4
1 September 2026 · build b2671
Mac · the accuracy pair, notarized
- Abandoned words are finished the way you meant them. When a word is cut off mid-way (“actiona… actionable”), the app completes it from context whenever most of the word was spoken — in both stutter-setting states. Only a small fragment is governed by the setting: ON (the default) removes it, OFF keeps it exactly as heard.
- A real word is never rewritten into a longer one. A fragment that already is a valid standalone word stays as that word — turning a correctly spoken word into a different, wrong word is the worse failure, so the app refuses to.
- Every completion is on the record. Each completed or removed fragment is filed for review, and the raw transcript always keeps what was actually said.
- Repairs can no longer be squeezed into silences too short to hold them. The refusal now happens before the text is pasted, using the same measured-gap law the after-paste review uses.
- Notarized by Apple. macOS opens this build without any warning.
10.4
1 September 2026 · build b2649
Mac · the dev season's accumulated fixes, in one notarized refresh
- Short recordings can no longer invent words. Very short clips used to decode empty or hallucinate; they now go through a padded decode, with a refusal guard on top for any text the audio does not actually contain.
- Review what the app found, completed. The self-found error list gains review-state pills (Reviewed / Partially reviewed / Reviewed‑no‑errors), sorting pills with live counts that double as filters, and findings that jump playback to just before the error — plus a fix for the journal not reloading fresh findings.
- Percent and number formatting hardened as whole classes. A spoken “ninety-nine point nine percent” can no longer render as a mangled range, decimals survive scale words, and formatting rules must now close an entire class of error — never just a list of examples.
- Email dictation welds correctly. A spoken name + “at” + domain joins into a proper address by shape, a swallowed “at” is reconstructed, and a learned habit can never rewrite a domain.
- Clipboard protection, round two. The clipboard lane ships its decided design, and History search now highlights what you searched for.
- Paste recovery's second act. The recovery card's ⌘V restores your text directly — flagged words and struck deletions in the karaoke view now jump the playhead when clicked.
- Measured, not guessed. The accuracy pipeline was replayed across 5,036 recorded utterances with zero unexplained movements — every accuracy claim in this span traces to a measured run.
- Notarized by Apple. macOS opens this build without any warning — the first notarized cut since the original 10.4 release.
10.4
25 August 2026 · build b1851
Mac · the self-fix season's newest corrections
- Numbers spoken while looking something up stay whole. “build. 1590” becomes “build 1590” — the thinking-pause period no longer splits a number from its noun.
- Seam phantoms collapse. Doubled half-words minted at long-audio seams (“go go-ahead” class) are recognized and removed; real stutters and deliberate repeats stay.
- Abbreviations no longer capitalize what follows. “Mr. was” keeps its lowercase — an abbreviation's period is not a sentence end.
- Every dictation and error report now carries its build number, so a report can always be matched to the exact software that produced it.
- Review what the app found. The self-found error list shows plain-English findings with one-tap Real / Not-an-error verdicts, your notes, and an “anything we missed?” field.
- No recording is ever discarded. Audio retention now covers the Chinese/Japanese/Korean engine path too — even a dictation that produces no text keeps its recording for recovery.
- Dev-signed build (not notarized) — macOS asks once to confirm opening.
10.4
22 August 2026 · build b1589
Mac · the same 10.4, quietly better — the self-fix season's corrections ship
- The worst hallucination class, gone at its cause. The microphone fade-in boost was measured to be net harmful — it is now off by default, ending the class of invented words it created.
- Phantom-word protection widened. The zero-trace check now guards words of every length at reading pauses — not just short ones.
- Email dictation repaired. Bare names weld correctly at mail providers, a swallowed “at” is reconstructed, and a learned habit can never rewrite a domain.
- One-spoken-value duplicates collapse. “Opus 5 V” and “91% percent” class echoes are recognized and removed with timestamp proof.
- The accent learner hardened. No more one-sighting pairs; common words are protected outright.
- Recovery you can trust. Paste-recovery cards are more accurate and permanently journaled.
10.4
17 August 2026
Mac · the accuracy season — invented words die, and every claim is measured
- Words you never spoke are refused. Dictation engines can quietly invent a small word at the seams of long audio. Now every word must leave a trace in the actual audio: a zero-trace sweep checks the dictation against what was heard before anything is pasted, whole phantom phrases are recognized and refused, and a lone spoken letter ("Whispr X") can no longer be swallowed by a longer word.
- Measured, not promised. The release was re-run across all six public benchmark sets against the previous public build: every single set improved, with roughly twenty genuine fixes for every new error introduced — and about 1,100 previously-dropped words restored to their dictations.
- Sound-alike echoes collapse with proof. When the engine writes one spoken thing twice in two forms ("for Four"), the echo is removed using the audio's own timestamps as evidence — deliberate repeats and real stutters are speech, and stay.
- Capitalization follows evidence. Mid-sentence capitals are now decided by how a word is actually used — in the language and in your dictations — instead of the engine's whim; names you've taught keep their capitals, and cold-start words are left exactly as heard.
- Teach it in seconds. Dictionary and voice-profile additions take effect in a running app within seconds — no restart — and your in-app Dictionary always wins any conflict.
- Hear it back in your own voice. Speakback can read your dictation aloud in a clone of your voice — built on your Mac after a one-time consent, never uploaded — with karaoke-style subtitles highlighted word by word, adjustable speed, and polite pausing whenever you start a new dictation.
- This download is the notarized 10.4 — Apple-signed, stapled, opens with no warnings.
10.3
13 August 2026
Mac · the self-learning release — the app now learns how you speak
- Your corrections teach it — once. Fix a misheard sound-alike word ("parody" when you meant "parity") and the pair is learned: from then on, wherever that mishearing appears, your own dictation history decides the right word. It never guesses — no evidence means the engine's word stands, numbers are never touched, and your Dictionary always outranks it.
- Your accent, understood. Sound-alike fixes route to your voice profile automatically and arm after a couple of sightings — and deliberate words are protected: say "a parody of…" and it stays exactly what you said.
- Word repair got a shape rule. The spelling healer can only fix a misheard non-word to a neighbor with the same spoken shape, and it will never "correct" a word you actually use — your technical vocabulary is safe by construction.
- Echoed acronyms collapse with proof. When the engine writes one spoken acronym twice in two spellings ("SDK STK"), the pair collapses to the one you use — while a deliberately repeated acronym is never touched.
- The correction watcher got faster and more honest. It checks for your fixes seconds after a paste, keeps watching patiently on long texts, and no longer gives up silently when macOS itself stops reporting focus — it waits for the system to heal instead.
- This download is the notarized 10.3 — Apple-signed, stapled, opens with no warnings.
10.2
12 August 2026
Mac · the big one — everything proven since 10.1 went public, in one download
- Long dictations now land almost instantly. The Mac transcribes in the background while you speak — releasing the key only finishes the last few seconds, so a ten-minute dictation pastes about as fast as a short one. The target app still receives your text once, at the end — no mid-speech rewriting on your screen.
- Power-cut and crash protection. While you dictate, everything needed to rebuild your words is continuously saved on your Mac. If the machine loses power or the app is killed mid-dictation, the next launch recovers it into History, honestly marked — proven with a real mid-dictation force-kill.
- The book-length accuracy round. An entire 3.4-hour audiobook was dictated through the app and scored word by word: the class of silent "holes" that long dictations could develop was root-caused and fixed, taking the book from 84.9% to 95.0% correct — and the speech engine itself was upgraded, killing several phantom-word classes at their source.
- Echoes refused. The engine sometimes writes one spoken thing twice in two forms ("take 18 eighteen seconds", a "10.2" growing a phantom "two"). Those echoes are now recognized and refused — while spoken years come out as years ("nineteen eighty-four" → 1984) and counting, lists, and repeated real words are never touched.
- The second listen. The seams where long-dictation audio windows meet are independently re-heard and the readings adjudicated, and weak-evidence words must be witnessed before they're kept — the app never pastes a guess it can't hear twice.
- Translate as you speak. Press fn+X and your dictation lands translated — using Apple's on-device translation with a local backup model, so nothing you say leaves the Mac. One undo chip if you want the original back.
- Mixed Chinese/English dictations now recognize and format each part under its own language's rules, and a dictated series of numbers renders in one consistent style.
- Fabled Flow can read back to you. Press fn+S and it reads the focused document, your selection, or pasted text aloud — with word-exact resume per document, a scrubber, and a follow-along display. (A full karaoke section built on the same machinery is in testing.)
- Sign in with Apple joined the email-code sign-in.
- The recovery card grew honest instincts. A card can never again stay silent over a possibly-lost paste without an actual verification look; the card's close button is measured against what's really drawn, so it always works; once you've moved on, a late card arrives as information (age-labeled) and can't re-paste by accident; and a card withdraws itself the moment your text visibly lands.
- Cold starts died. Waking or unlocking your Mac re-warms the speech engines before your first press, and capture always starts from the keypress itself — the first dictation after a long idle is no longer the slow one. (The microphone still only opens per dictation — the privacy promise is untouched.)
- Install integrity. Only one Fabled Flow can ever run — a broken or stray copy refuses to launch and cleans up after itself; permission entries that macOS quietly breaks (the "it asks for Accessibility that's already ON" state) are detected and healed with one explained dialog; and a full disk can no longer take the app down while it writes its own diagnostics — it announces, never aborts.
- This download is the notarized 10.2 — Apple-signed, stapled, opens with no warnings.
Fabled Flow