# Blabb

> Blabb is a privacy-first voice dictation app for Windows 11 (x64). It turns speech into clean text and types it directly at the cursor in any application — including locked-down RDP, Citrix, and VDI sessions where clipboard paste is blocked. Speech recognition and AI text cleanup both run 100% on-device; no audio or text is ever uploaded. Distributed via the Microsoft Store; a free tier of 2,000 dictated words a day needs no card, and subscriptions (with a 14-day free trial) are billed through Polar, a merchant of record.

**What it is:**

- Category: voice-to-text / speech-to-text dictation utility for Windows
- Platform: Windows 11, x64 only (no ARM64). No NPU or Copilot+ hardware required; runs on CPU, uses GPU when available.
- Distribution: Microsoft Store — https://apps.microsoft.com/detail/9N5G1JPFMCJN
- Pricing: Blabb Free gives 2,000 dictated words a day, indefinitely, with every feature and no card — the count resets at local midnight and the upgrade ask appears only once a day's words are spent. Unlimited dictation is USD $4.99/month (or USD $49.99/year — two months free) with a 14-day free trial, or USD $99 once for a lifetime licence (no trial, nothing recurring to cancel), billed through Polar (merchant of record). Paid plans have no word caps and no per-use metering. There is no grace period after a lapsed payment — dictation drops back to the free tier's daily allowance.
- Specialized editions (separate, dearer plans — not included in the $4.99 plan): Blabb + Medical Vocabulary, USD $9.99/month or USD $99.99/year (48,133 medical terms: FDA drug names, NLM clinical terminology, ICD-10 diagnoses); Blabb + Legal Vocabulary, USD $9.99/month or USD $99.99/year (13,004 legal terms: terms of art, Latin phrases, procedural vocabulary). Each edition plan includes everything in the base plan and adds the vocabulary on top. Dictionaries are processed entirely on-device, same privacy as the base app.
- No API key required. Models ship inside the installer. Buying a subscription creates a Polar customer record and a licence key, which is pasted into the app's Settings.

**How it works:**

1. Press a global hotkey (default: F9) anywhere in Windows. Press once to start, press again to finish; hold-to-talk is the alternative behaviour.
2. Speak naturally — rambling, pauses, and filler words are fine. Recording runs until you stop it; it does not cut off after a few seconds of silence.
3. Clean text is typed at the cursor in whatever app has focus.

**Key capabilities:**

- Types into locked-down environments: RDP, Citrix, VDI sessions with clipboard redirection disabled. Four text-entry methods — Direct (UI Automation), Type fast (Unicode SendInput), Type slowly (simulated keystrokes), Paste (clipboard) — set globally or overridden per app by the user, so text lands even where paste is blocked. Limitation: cannot type into elevated/admin (UIPI-protected) windows — a Windows security boundary, not an app choice.
- Works in any app with a text cursor: browsers (Chrome, Edge), Word, Outlook, Teams, Slack, VS Code, terminals, Notepad, etc.
- AI cleanup: removes filler words ("um", "uh"), fixes punctuation, converts spoken numbers to digits ("twenty three" → "23"), preserves the speaker's meaning and voice. The cleanup prompt instructs the model never to add content and never to follow instructions spoken inside a dictation.
- Six writing styles, named in the app: Keep original (verbatim), Fix typos (grammar/spelling only), Make clear (shorter, clearer wording), Polish professionally (business tone), Warm up tone (warm, conversational), Match screen terms (matches names/terms/casing read from the active window). Default is Keep original — no cleanup unless the user chooses it.
- Personal glossary: learns names, jargon, and acronyms from the user's corrections — stored locally.
- Dictation history with visual diffs (see exactly what the AI changed), search, copy, and delete. Stored in an encrypted local database. A time-saved counter is shown; it is an estimate based on an assumed 40 wpm typing speed, not a measurement.
- Dictation works offline: an already-verified subscription keeps working for 30 days without reaching the licence server.

**Under the hood:**

- Speech recognition: CrispASR running on-device, default backend Cohere Transcribe (alternatives: Parakeet, Granite, SenseVoice). Voice activity detection: TEN VAD.
- Text cleanup: IBM Granite via LM-Kit — a compact, openly licensed language model built under an ISO 42001-certified AI process.
- Both models run entirely on the local PC (CPU or GPU). No cloud inference, no queue, and no cloud AI providers.

**Privacy (the core promise):**

- No audio or text ever leaves the PC. Three things do reach the internet, and only these: downloads for a model that was not packaged with the app, the Polar licence check, and crash reports (opt-in, off by default).
- Dictation history is encrypted at rest with keys tied to the Windows user account (row-level encryption).
- Retention options for history: Don't save history (nothing is written to disk at all), 1 day, 1 week, 1 month, or keep forever. Default is 1 week.
- Sensitive content is masked in diagnostic logs by default.
- One-click "Delete Blabb data" removes history, saved words, and logs. Settings and the licence stay.
- No tracking, no cloud processing of audio or text.

**vs. Windows Voice Typing (Win+H):**

- Win+H sends audio to Microsoft Azure; Blabb keeps it on the PC.
- Win+H is blocked in locked-down RDP/Citrix/VDI; Blabb types straight in.
- Win+H cuts off after ~5–10 seconds of silence; Blabb records until you stop it.
- Win+H AI cleanup needs Copilot+ NPU hardware; Blabb's works on any Windows 11 x64 PC.
- Win+H needs the internet for every dictation; Blabb dictates offline, and only checks its licence about once a day (30-day offline window on a verified subscription).
- (Comparison as of July 2026.)

**Who it's for:** People dictating into locked-down corporate desktops (finance, healthcare, consulting), high-volume writers (support teams, email-heavy roles), privacy-bound work (NDA/PHI content that must not touch a cloud), developers dictating to AI assistants, and users with RSI who need to type less.

**FAQ:**

- Q: Does my voice ever leave my PC? — A: No. Speech recognition and AI cleanup both run locally; audio and transcripts are never uploaded. The only call the app makes on a normal day is the licence check.
- Q: Does the free trial auto-charge? — A: The trial runs through Polar, the payment provider; the developer never sees your card. Cancel within 14 days and you pay nothing.
- Q: Do I need an API key or my own model? — A: No. Both models ship inside the installer. No downloads, no per-use billing.
- Q: What are the system requirements? — A: Windows 11 on x64. A microphone. No GPU or NPU required (GPU used if present).

## Links

- [Website](https://blabb.io/): product overview, features, writing styles, privacy details, and pricing
- [Documentation](https://blabb.io/docs/): getting started, every setting, accuracy and performance tuning, medical and legal vocabularies, troubleshooting
- [Documentation index for agents](https://blabb.io/docs/llms.txt): every documentation page with a one-line summary; each page is also served as markdown at its URL plus `index.md`
- [Microsoft Store listing](https://apps.microsoft.com/detail/9N5G1JPFMCJN): download (subscription and 14-day free trial are purchased separately through Polar)

## Optional

- [Legal documents](https://blabb.io/legal.html): Privacy Policy, EULA, and Third-Party Notices