A macOS menu-bar voice input. Hold the hotkey, speak — text streams into your cursor in real time, and every sentence sediments into a searchable, private memory layer that lives only on your Mac.
The iPhone version is a standalone system keyboard focused on voice input across apps. Its features and workflow are not identical to the Mac version.
Local engines and Bring-Your-Own-Key paths are free forever, no subscription. Pro removes all cloud quotas with zero configuration.
Two on-device recognition engines (SenseVoice / Apple Speech Recognition). Nothing leaves your Mac.
Plug in any OpenAI-compatible API key (DeepSeek / Kimi / OpenAI / your local OpenAI-compatible server). Unlimited tidy, you control the cost.
Zero-config cloud ASR + cloud AI tidy. Removes the 60 min/mo + 50 tidy/day quota. For people who just want it to work.
Input on top, voice archive in the middle, AI memory underneath.
Hold, speak, release. Mixed-language, homophones, fillers handled silently. Under 1.4s.
Every line archives locally with source app, time, tags. Search, filter, export.
7 personas review your week. A weekly MBTI sketch. 3–5 quotes worth echoing.
7 built-in personas plus your own. Contrast itself is the memory.
Introspection up, decisions slower. Three returns to the "sunk cost" theme.
"Speed is iron law — no feature may add perceived latency."
Casing, punctuation, and unit spacing — handled locally in <5ms, no LLM call.
LLMs handle semantic judgment (homophones, fillers). The local engine handles format (brand casing, punctuation spacing, unit spacing). Two layers; both wins.
45+ brand names auto-corrected. Spacing after English commas/periods. Half-width parentheses kept for ASCII. Every rule toggleable.
Pin source app at record-start. Switch windows — still lands right, with clipboard fallback.
Stop talking, wait 1–2 seconds, polished text lands in one shot. You never see the raw.
Local pinyin injected into prompt. Homophone pairs no longer confused.
Your edits on AI output extract as candidate rules. Accept from menu bar.
AI models, dev tools, Apple products built-in. "Cursor" stays "Cursor".
More text, more transparent. Breathing glow hints AI is working. Original never disappears.
Audio, text and history all live on your Mac. One single number leaves — seconds per recording.
Audio and text land in the app's own directory, auto-backed up on launch. Uninstall takes it all with you.
Each recording sends only its length to the global pulse. No identity, IP, content, or context. Toggle off in Settings.
API keys sit in macOS Keychain, never on our servers. ASR runs directly against Volcengine, nothing persisted.
We don't pile features — we only ship what we believe earns its place.
With “AI deep correction” on, misheard words in Direct Insert are fixed in your text field about a second after you stop — and nothing is touched if you start typing right away. The menu bar “Last phrase” line brings back Copy, Re-transcribe and Re-tidy, and the recognition engine can be switched right from the menu bar. Details the redesign had hidden are back, and the English interface is fully translated.
All-new interface: six clear sidebar entries, an overview that first tells you whether you can speak and which key to hold, scannable text lists for records and flash notes, settings without stacked cards, light and dark following the system. The recording overlay drops constant animations and glows, appears the instant you press and uses far less CPU; the global hotkey is no longer affected by UI stalls, and large histories open noticeably faster. Local recognition remembers names and phrasings you add to “My words”, and text appears progressively while you speak. A new “Usage & performance data” switch can be turned off anytime.
Local recognition now waits for the full result, so long recordings are no longer cut off or replaced by preview text; quiet short clips are no longer dropped as "no sound"; rapid back-to-back dictation no longer disrupts the previous clip. The recognizer stays resident for faster startup and corrupted model files repair themselves. Expired AI account keys fall back to another tier with a flag in Settings, and crashes are reported automatically.
Local recognition is faster and more accurate, and long recordings no longer slow down over time. Fixed occasional missing words at the start, text disappearing after the "Tidying" indicator, and the model name disappearing after undoing AI Tidy. AI Tidy and usage stats are now more accurate, and onboarding includes an optional discovery question. The interface is available in Traditional Chinese, Japanese, Korean, German, French, and Spanish, with one-line terminal installation.
Cloud AI cleanup is much faster — results appear almost as soon as you finish speaking (especially on networks in China), and the first cleanup after launch is faster too. Fully automatic updates: new versions download and install silently while idle. A batch of Chinese corrections and brand-name fixes never actually matched before and now work as intended; translation mode no longer rewrites English words by mistake. Settings are much simpler, and you can see how many fixes the rules made for you this month. Account and membership merged into one card — sign in with your purchase email to restore Pro. Plus one-command Homebrew install.
Smoother onboarding: after granting permissions, one big button restarts the app so they take effect immediately; the onboarding window pops up more reliably on first launch. Launch at login is now on by default, so VoiceInput is still there after you reboot. Local direct-insert is smoother — text appears more continuously as you speak.
Cloud recognition / AI cleanup is faster: connection reuse makes text appear sooner after release in most cases. Translation / bilingual is more accurate: proper nouns (company / product names) are recognized better.
No. The free plan works out of the box: local engines are unlimited and VoiceInput Cloud recognition gives 60 minutes a month. For AI Tidy with no quota, add your own API key (DeepSeek, Kimi, OpenAI or any OpenAI-compatible service); you pay your provider directly.
VoiceInput is macOS only: hold to talk, local engines work fully offline, Chinese-English typography runs on your Mac, and what you said is stored locally and searchable. See the VoiceInput vs Typeless page for a side-by-side with sources.
macOS 14.0+, Apple Silicon + Intel. 25.5 MB DMG, non-App-Store, Sparkle auto-update.
Microphone, Input Monitoring, Accessibility. Granted once via the onboarding page.
No. AI Tidy does three things: fixes homophones, drops fillers and adds punctuation. It does not add what you did not say, and the default Tidy style does not translate. To insert text as-is, use the Direct Insert key (Right Option by default), which skips AI.
Yes. Export TXT, CSV or JSON from Settings. Copy the database file to the same path on a new Mac to move it.
Yes. Clear API config, all memory features stop. Local typography engine keeps running.
Honest, side-by-side comparisons against the tools you're probably also evaluating.
Faster on Chinese, plus a memory layer Superwhisper doesn't have. Read the side-by-side →
Different categories: Wispr rewrites tone, VoiceInput keeps memory. Compare →
Hold-to-talk, AI cleanup of fillers, Chinese-English typography, searchable history. Side by side with Apple Dictation →
Download · grant three permissions · hold right Option. Thirty seconds to get it.
Download VoiceInput_v0.91.0.dmg · 25.5 MBbrew install --cask voiceinput-app/tap/voiceinput