v1.8.4 — Keep trying until your internet comes back

• Recordings that can't be transcribed now wait safely on the device and resend themselves automatically the moment a working internet connection returns — no more giving up after a few tries
• Waiting recordings survive closing the keyboard, switching apps, and even restarting the phone or the app; nothing is deleted until it has actually been sent
• The keyboard now checks for a real connection before sending and no longer hangs for minutes trying to reach a dead network

v1.8.3 — Don't lose recordings when the network drops

• Failed transcriptions are now retried automatically (3 attempts with a short backoff), so a brief Wi-Fi↔mobile hand-off no longer eats your dictation
• If a recording still can't be sent, it's kept instead of discarded: a red counter and "resend" button appear under the backspace key — tap it once the connection is back to transcribe everything that was waiting
• Permanent errors (wrong API key) stop retrying immediately but the recording is still kept, so you can fix the key and resend

v1.8.2 — Strip trailing "Субтитры создавал X" hallucinations

• Removes Whisper end-card credits appended after legitimate speech (the classic "Субтитры создавал DimaTorzok" tail, plus "Subtitles by …", "Sous-titres réalisés par …", "Untertitelung im Auftrag des …", "Спасибо за просмотр", etc.)
• Real sentences that just happen to mention subtitles ("Я смотрел фильм. Субтитры были на английском.", "Включи субтитры пожалуйста.") pass through unchanged
• Initials in credit lines ("И. Иванов", "А. Сёмкин") are recognized as part of the credit, not a sentence boundary

v1.8.1 — Strip trailing notes from post-processing output

• Removes trailing "Note:" / "Footnote:" / "Disclaimer:" blocks (in any supported language) that post-processing sometimes appended after the actual text
• Removes stray separator lines (---, ***, ___) at the end of the output
• Strengthened the post-processing prompt to forbid these in the first place; the strip step is a safety net

v1.8.0 — Stop names list from leaking on silence

• Detects when Whisper returns the vocabulary list as a "transcription" of empty recordings, and drops it
• Catches more silence artifacts: subtitle credits with "Корректор/Редактор", thanks lines in German/French
• Stops eating real one-word dictation: short "Привет" / "Спасибо" / "Хорошо" pass through correctly
• Three-name dictations like "Альфия, Алсу, Рустем" are preserved while seven-name hallucinations are not

v1.7.9 — Better silence handling

• Catches more empty-recording artifacts: "Спасибо", "Vielen Dank", subtitle credit lines
• Stops echoes of the style hint from leaking into the field
• Short clips no longer get silently dropped — only long silent recordings are filtered

v1.7.8 — Fix vocabulary leak on silence

• Fixed: when nothing was said, post-processing could output a list of vocabulary words
• The vocabulary is now passed as a strict spelling reference, never as content to insert

v1.7.7 — Vocabulary in post-processing

• The custom vocabulary list is now also sent to the post-processing model (Fix, Shorten, Emoji, Rhyme, Translate) — rare names survive cleanup instead of being "corrected" to common spellings
• Terminal mode is excluded by design

v1.7.6 — Custom vocabulary

• New "Custom vocabulary" screen in settings — add names, technical terms, and other words that get misrecognized; the list is sent as a hint to the recognizer to improve accuracy
• New "Drop period for single-word output" toggle — useful for voice search in contacts or search boxes, where a trailing period is unwanted
• Both features can be edited at any time from Settings

v1.7.5 — Smarter text cleanup

• Fix and Shorten modes now handle self-corrections ("wait no", "не, я имел в виду") — only the corrected version is kept
• Better removal of false starts, stutters, and accidental repetitions
• Preserves technical terms, proper nouns, and the speaker's voice more reliably
• Empty or filler-only input now produces no output