- Setup no longer gets stuck on "Input Monitoring." On some Macs that step could open an empty settings pane with nothing to switch on, trapping you before you could finish — even though Lispr's dictation works fine without it. The step is now optional and registers Lispr correctly, so onboarding always completes.
Changelog
What's new in Lispr.
Lispr ships fast — sometimes more than once a day. Below is what changed in every release since early access opened.
Every install also auto-updates via Sparkle from the appcast feed.
- The mic recovers from more stubborn failures. On a few Macs the microphone could get stuck in an error state and refuse to start, even after a retry — Lispr now rebuilds its audio session from scratch to bring it back, instead of leaving you with a dead mic.
- No more slow first dictation after a pause. When Lispr had been idle for a while, the next dictation could stall for several seconds on a stale network connection. Lispr now refreshes the connection after an idle stretch, so dictation stays fast even after a break.
- Fixed a freeze when opening History. On some Macs, opening the History window could peg the app and freeze it — Lispr now renders the list far more efficiently, so History opens instantly and stays smooth.
- Confirming the mic recovery actually helps. Behind the scenes, Lispr now notes when its automatic recovery quietly rescued a dictation that would otherwise have been lost — so we can see the fix from 0.78 working for you. Nothing changes in how you dictate.
- Slow-to-wake mics no longer cost you a dictation. When the microphone needs an extra moment to start, Lispr now briefly waits for the first audio instead of throwing the recording away — so quick or just-started dictations still come through. It also re-warms the mic after your Mac wakes from sleep, so the first dictation isn't the one that gets lost.
- Mic that wouldn't start now heals itself. Fixed cases on some Macs where a dictation began but captured no audio — Lispr now detects when the microphone was interrupted or never armed and automatically re-arms it for your next recording.
- Paste now lands in Telegram, and other Qt/Electron apps. Fixed a case where dictated text reached your History but silently failed to insert into the focused field — Lispr now presses the real Command key for paste, so the text shows up where you typed.
- Auto-stop on silence. Forget to end a dictation? Lispr now stops on its own after a stretch of quiet. Pick the timeout in Settings — Off, 30s, 1 min, or 2 min — now localized in all 35 languages.
- Smarter hallucination filter. A quiet or silent tap no longer turns into an invented phrase like "Thank you" or stray subtitle credits — Lispr now lines up the transcript against where it actually heard your voice and drops the rest, in any language. Real speech, even a faint whisper, is kept.
- Your audio survives a crash. Recordings are now autosaved as you speak and recoverable from History after a force-quit or crash, and long dictations no longer get cut off — the recording limit went from 10 minutes to 2 hours.
- Small maintenance release to sharpen how we measure silence-vs-speech in the field, so the upcoming auto-stop tuning lands right.
- Under-the-hood tuning for the silence detector. Groundwork toward smarter "is this real speech or just background noise" decisions — no behavior change yet.
- Behind-the-scenes reliability and site improvements. No changes to dictation behavior.
- Fixed a startup crash on some Intel Macs. A 0.70 change could crash the app on launch on certain Intel machines — this update resolves it. If you were affected, grab the latest build from lispr.ai.
- Internal tuning for the on-device voice-detection work. No user-facing changes.
- Smarter silence detection, on-device. Groundwork for telling real speech from quiet/noise entirely on your Mac — the foundation for killing the "Thank you" that silent clips sometimes invented. Measuring only in this build; no change to what gets inserted yet.
- A dead mic now tells you. If your microphone never actually turned on, you get a clear "Mic didn't turn on — try again" instead of a silently empty text field — no more lost dictations you can't explain.
- Calmer update reminders. When a release fixes a known issue, a quiet in-app note lets you know — no version-shaming, no pop-ups mid-dictation.
- Hardened update integrity. Updates are now verified server-side before they reach you, so a download that gets corrupted in transit retries automatically instead of failing with an "improperly signed" error.
- Smarter speed tracking. Lispr can now measure the true time from key-up to text on screen so we can chase down real-world lag — numbers only, sent after your text is in, with zero impact on dictation speed. Honors your Diagnostics setting.
- One clear privacy switch. The old "Send on errors" toggle is now "Diagnostics & usage stats" and controls all anonymous technical data — turn it off and nothing technical ever leaves your Mac.
- Plainer Settings. The two data controls now sit side by side, and the privacy policy spells out exactly what each one covers.
- Opt in to help improve Lispr. A new Settings switch lets you share recordings so we can make dictation more accurate. It's off by default, and you decide after reading exactly what's shared.
- Nothing leaves without your say-so. The contribution is fully optional and independent from normal dictation — your audio and text stay private unless you turn it on.
- Fresh build for everyone. Rolls the recent language-accuracy and Concise formatting fixes out to all users via auto-update.
- Fixed Russian getting turned into Ukrainian. The auto-language feature from 0.60 could drift and force the wrong language on some users, rewriting what they actually said — it's now disabled and any affected profiles are reset on first launch. You can still lock a language by hand from the menu.
- Concise formatting actually condenses now. On polite, hedge-heavy speech Concise mode was leaving everything as one block — it now structures and trims reliably while keeping your tone and meaning intact.
- Clean up your dictation as you speak. A new Formatting menu adds two optional modes — Light breaks your speech into paragraphs and lists without dropping a word, and Concise also trims fillers and stutters. Default is Off, so nothing changes unless you turn it on.
- More reliable mic-gain help. Fixed a case where the "raise your input volume" offer never actually appeared for the people who needed it most.
- Pick your dictation language. A new menu lets you lock the language Lispr transcribes in — fixes short Ukrainian and Belarusian phrases that were getting mistaken for other languages.
- It learns your usual language. Lispr quietly remembers what you dictate in and leans that way over time, with a source-language badge shown in the pill.
- Quiet voices no longer get dropped. The silence check now looks at your speech peaks and signal-to-noise, so soft or distant speech gets transcribed instead of being thrown away as "no audio."
- One-click mic fix. When your input volume is too low to hear you, Lispr now offers to raise it for you — and catches dead AirPods or external mics early so a long recording isn't wasted.
- Fewer false "no audio detected" banners — a one-off blip stays quiet and just retries instead of nagging you.
- Open History from anywhere. Press
⌃⌘Yin any app to pull up your dictation history without switching back to Lispr. - Fixed empty recordings on macOS 12 Monterey. A codec quirk was silently truncating audio on Monterey so dictations came back blank — now they transcribe normally.
- Localization polish for Russian and Korean wording.
- Groundwork release. Mostly under-the-hood work to track down a rare "recording too short" failure, plus localization polish. No behavior change for most users.
- The mic helper now diagnoses more cases. It can spot a dead mic with no signal, an input-volume slider turned down, voice-processing software flattening your audio, or a room too noisy to transcribe — each with targeted advice or a one-tap jump to the right Settings.
- All of the mic-helper guidance is now hand-translated into every one of Lispr's 35 languages.
- Lispr now tells you why it can't hear you. After a couple of silent recordings in a row, a banner appears under the pill with the likely cause — and a one-tap fix.
- If the mic you picked has disconnected (AirPods undocked, USB unplugged), the banner offers to switch you back to the built-in mic. If microphone access is off, it takes you straight to the right Settings pane.
- Sharper silent-recording diagnostics — Lispr can now tell when the built-in mic is muffled because the lid is closed in clamshell mode. Also fixes a build issue affecting the macOS 11 target.
- Translation learns from what you actually said. When you dictate-and-translate, Lispr now picks up your custom vocabulary from the original speech, not the translated output — so everyday target-language words stop getting mistaken for your brand terms.
- History keeps the original. Translated dictations show the translation with a small
→CODEbadge, while the text you spoke is preserved underneath.
- The live usage counter stops flaking. The stats endpoint behind the counter on lispr.ai was throwing a 504 on roughly a third of loads — it now talks straight to
proxy.lispr.aiand loads reliably.
- Broken-mic visibility. When a recording comes back empty or silent despite you speaking, Lispr now captures the mic, level, and permission state so a dead or wrong input device can be tracked down instead of just failing quietly.
- Instant translation — speak one language, type another. Dictate as usual, and any time mid-sentence press a second key: Lispr inserts the translation instead of the verbatim text. Hold ⌃ for one target language, ⇧ for another. The recording pill shows a little →EN badge so you always know what you’ll get. Let go a touch early and it still counts.
- Two configurable slots, 32 languages. Pick the key and the target language for each slot in the new Translate window. Sensible defaults are chosen from your Mac’s language on first launch. If the translation ever fails, Lispr falls back to inserting the plain transcript — you never lose what you said.
- Lispr learns your brand names — automatically. The new Vocabulary feature watches how Whisper tends to mishear words you actually use — "Котбридж" instead of Codebridge, "Монобанк" instead of Monobank — and quietly feeds the correct spellings back to the model. Enabled by default so you get the accuracy boost without finding a hidden switch. Open Vocabulary from the menu to see what got learned, edit anything, or turn it off.
- The menu got a major cleanup. Top-level rows dropped from 17 to 9. Trigger key and trigger side merged into one Trigger submenu. Rare items (Launch at Login, Diagnostics, Setup Guide, Accessibility) moved into a single Settings submenu. About slid to the bottom to match how macOS menu-bar apps actually work. Check for Updates is now an Option-alt of About — hold ⌥ to reveal it.
- Pick which side of the trigger key counts. If you use option+e for accents in PT/FR/ES/DE/IT, you can now restrict dictation to Right Option only and keep left for typography. Per-key memory — Option's preference doesn't affect Shift's. Available in onboarding and in the menu under Trigger.
- The menu hint follows your trigger. If you picked Control, the hint now says "Hold ⌃ to speak" — not the hardcoded ⌥ glyph that used to lie regardless of your choice. Same for onboarding copy and trigger names across all 35 languages.
- History and Vocabulary now feel like one app. Shared chrome, typography, spacing. Both use the system window background — which means History finally works in dark mode (it was hardcoded to a cream background for six months).
- Whisper hint is no longer Russian-biased. The vocabulary went into Whisper as
"Имена: Codebridge, Monobank."with a Cyrillic prefix word. On English/Portuguese/Indonesian audio that word skewed Whisper's language classifier toward Cyrillic output. It's now a bare comma list — same accuracy boost for proper nouns, no language bias. - Vocabulary UI got quieter. Removed the colored "confidence dots" and the count column — they didn't drive any user action. Edit and delete are now front-and-center.
- Auto-diagnostics for the vocabulary feature. When vocabulary breaks for a real user (file corruption, save failure, server unreachable for three retries), we now find out — so problems get fixed before they spread.
- Lispr learns your brand names. The new Vocabulary feature watches how Whisper tends to mishear words you actually use — "Котбридж" instead of Codebridge, "Монобанк" instead of Monobank — and quietly feeds the correct spellings back to the model. Open Vocabulary from the menu to see what got learned, edit anything, or remove a term.
- The menu got a major cleanup. Top-level rows dropped from 17 to 9. Trigger key and trigger side merged into one Trigger submenu, rare items moved into a single Settings submenu, and About slid to the bottom to match how macOS menu-bar apps actually work.
- Pick which side of the trigger key counts, and the hint follows your choice. Restrict dictation to Right Option only and keep Left for typing accents in PT/FR/ES/DE/IT — remembered per key. The menu hint now reads "Hold ⌃ to speak" for your actual trigger instead of a hardcoded ⌥, in all 35 languages. History also gains proper dark mode.
- Follow-up fix for the 0.44 rollout — resolves a brief hiccup that could make the download page momentarily unavailable. Plus quieter under-the-hood timing instrumentation to chase down the last rare slow-dictation reports.
- No more stuck recording pill on flaky networks. A 30-second watchdog now cancels a stalled upload instead of letting the pill hang forever — the culprit was a stale pooled connection on spotty mobile Wi-Fi. Lispr also refreshes its connection the moment your Mac wakes, so the first dictation after sleep doesn't stall.
- Smarter routing to the nearest edge. Lispr now picks the server by the data center actually handling your traffic, not your account country — which fixes an odd Sydney detour some users hit on hotel and captive Wi-Fi.
- Lispr re-warms its connection to the proxy as soon as your Mac wakes up — the first morning dictation no longer pays a TLS handshake.
- Connection reuse across dictations. One shared TLS connection survives every dictation in a session — eliminates the 2–3 second lags some users hit on flaky Wi-Fi when each upload had to re-handshake.
- CRC self-test on the Opus framer expanded from 1 to 5 vectors — defends against a class of bug that previously crashed 0.41 on its first launch.
- Streaming Opus uploads + geo-aware relay. Keyup-to-text median dropped to ~300 ms (down from 1.5–2.5 s on most flows).
- Unknown URLs on lispr.ai now return a real 404 instead of silently serving the home page.
- Diagnostic
X-Lispr-*response headers — Lispr can be inspected remotely without storing logs server-side.
- Short and empty recordings now filtered before going to transcription — fixes about 9 out of 10 spurious "we didn't catch that" messages.
- The DMG and Sparkle appcast are served over a TCP-only proxy (
proxy.lispr.ai) — fixes auto-updates behind some restrictive networks.
- Esc cancels a recording. Press it at any time to drop what you've said before it gets inserted.
- Whisper "hallucination" filter — silent recordings no longer turn into invented sentences.
- Soft cap on recording length (10 minutes) with a localized warning — prevents accidental marathon recordings.
- Asia-Pacific traffic routed through a Sydney relay — measurably faster than the US path for the region.
- Direct Groq call by default; US Durable Object is the fallback for geo-blocked routes.
- Custom proxy domain
proxy.lispr.aion a TCP-only zone — bypasses HTTP/3 issues some networks have. - Mobile landing polish: footer padding, primary button alignment, language menu close behavior.
- No more extra leading space when continuing a dictation — Lispr only auto-joins after its own previous insertion.
- Transcription timeout pulled from 30 s down to 10 s — faster recovery when the network blips.
- Lispr now registers in System Settings → Privacy & Security → Input Monitoring properly (via
IOHIDRequestAccess) — onboarding flow is cleaner. - Smart Groq key rotation with an alert when keys exhaust — no more silent quota-out failures.
- Bundles
libswift_Concurrency.dylibin the app — Lispr launches cleanly on macOS 11 Big Sur. - The download URL is now served with
Cache-Control: no-store— a re-download always pulls the latest signed build.
- Auto-retry on transient transcription failures.
- Retry a failed dictation from the history view.
- Smart sentence joining. A period and a space are inserted automatically between consecutive dictations — no more running-on sentences.
- The recording pill in the menu bar now stretches and shrinks smoothly between phases (listening · transcribing · done) instead of snapping.
- Gentle "didn't catch that — try again" message on Groq 400 errors, instead of a silent failure.
- Live counter on the landing updates in real time (was a snapshot before).
- Time-saved counter now climbs the full unit ladder: minutes → hours → days → years → … → millennia.
- Double-tap latches the recording. Tap once and hold to dictate, or double-tap to latch — Lispr keeps recording hands-free until you tap again.
- Fn key as the push-to-talk trigger. Use Fn instead of (or in addition to) right ⌥ — handy if your Option key is already busy.
- Menu-bar icon no longer disappears on macOS 11–12.
- Onboarding spark drawn as a shape — no longer relies on the SF Symbol "sparkle" that's missing on older macOS.
- Minimum macOS pulled from 14 down to 11. Lispr now runs on every Big Sur and newer — full Apple Silicon and Intel coverage.
0.2 – 0.23 · Early access launch
17–18 May 2026
The first 22 releases — initial public availability. Push-to-talk via right ⌥, Whisper transcription via Groq on an edge network, ~4 MB menu-bar app, ~99 languages auto-detected, microphone picker, notarized DMG with Sparkle auto-updates, and the live usage counter on the landing.