Veylo preview data handling
- Dictation is explicit: the microphone opens when you start a session and closes on stop/cancel/limit. Audio is processed in memory. No automatic background recording is implemented. - Audio and enabled vocabulary are sent over HTTPS or encrypted WebSocket to Google's Gemini API. Veylo uses your API key and account quota. Google's terms and data policies govern remote processing; Veylo cannot promise zero retention by Google. Live transcription is an explicit preference, off by default: it sends audio while recording is active. Batch mode starts uploading after Stop. - Starting in 0.8.4, normal Auto/English/Hindi dictation uses the dedicated Gemini Transcribe model through the Interactions API. Requests explicitly set `store:false`, send bounded audio inline, and create no Files API resource or conversation chain. This opts out of interaction storage; Google's separate processing and account policies still apply. Roman Hinglish retains the general model for Latin-script output. - The explicit microphone test records locally for five seconds and discards its audio without a provider request. The connection test sends a model metadata request or opens the selected live model's setup with the key and configured vocabulary, without audio. Batch long dictation sends sequential audio segments with up to three concurrent requests. Live mode renews connections between acknowledged phrases before its session limit. These operations use your account quota. - Veylo does not send screenshots, surrounding text, telemetry or logs. Optional Windows target checks inspect field identity, editability, password status and selection locally before insertion. In explicit command mode, text transformations send the selected text (up to 16 KB) and your editing instruction to Gemini. The surrounding field value (up to 64 KB) is inspected locally to verify replacement and retained briefly in memory for undo; it is not included in the request. App and snippet commands resolve locally after recognition. Typed commands need no audio. Command previews, original selections and undo records are not saved in transcript history and expire five minutes after capture (two minutes for undo), or on dismissal/new capture/exit. Model output is text only and cannot launch tools. - Keys saved by the app are protected with Windows DPAPI in `%LOCALAPPDATA%\Veylo\v4\gemini.secret`. Importing an old key leaves `%LOCALAPPDATA%\Veylo\gemini.key` intact as requested. That legacy file is plaintext and remains managed by its owner. - Transcript history is off by default. When enabled it is stored in local SQLite, not encrypted by Veylo, with the selected retention period and a maximum of 2,000 entries. Windows account/disk protections apply. Turning saving off does not erase previous entries; use Clear history to remove them. - The current transcript remains in memory for review. A recording stays in memory during recognition and for retry for up to 120 seconds after failure; success, cancel, a new session or expiry discards the retry buffer. No retry audio is written to disk. Both original and processed text are saved when history is enabled. Explicit spelling corrections and snippets run locally. - Copy is explicit and places text on the Windows clipboard, where other applications or clipboard history may have access. Automatic insertion uses Unicode input, not a clipboard replacement. - The application makes no medical/legal privacy certification claim. Do not equate local history controls with Google's processing policy.
No API keys, private profile data, recordings or model weights belong in release packages.
Live mode retains complete audio in memory and automatically retries through batch recognition after Stop if the stream disconnects, falls behind or cannot confirm a phrase. Version 0.8 recovers audio after the last explicitly confirmed phrase using exact normalized sample offsets; if no phrase is confirmed, it retries the full recording. Confirmed text is joined once to the successfully recognized tail. Unconfirmed hypotheses are kept separately in memory as an explicitly unverified recovery draft; they never become a completed result automatically. Recovery results require review; failed recovery retains the full original WAV for a manual retry. This can resend previously uploaded audio and incur additional quota. Roman Hinglish uses batch recognition to retain the requested Latin script. Optional Smart live formatting changes the provider transcript; a recovered result can contain a Smart prefix and faithful batch suffix. The stored original is provider output, not a separate raw transcript.
Optional **Keep connection ready** retains one successfully completed provider connection for at most 20 seconds, or until the existing connection's age limit. It is off by default. The microphone remains closed between recordings; no synthetic audio or periodic keepalive audio is sent. A ping/pong check verifies the connection before reuse. The provider session remains open briefly and may use quota. Changing preferences, credentials, cancellation, failures, local microphone testing and app shutdown discard idle connections. Recognition settings and credential identity must match for reuse. The app does not preconnect at startup.
Provisional text is never automatically saved, inserted, or executed. On failure it remains in an explicitly unverified recovery draft with a manual Copy action; cancel, a new recording, successful recovery, microphone test, or exit clears that draft. An interrupted microphone capture preserves the available audio prefix in memory for recognition and labels the result incomplete. Interrupted results are not automatically inserted or saved to history, and interrupted commands are not executed. The usual retry expiry applies. All phrases must be acknowledged before a complete live result can be saved or inserted. The dedicated speech model returns completed short transcripts; long clips show completed segments. Roman Hinglish can show streamed responses after Stop. History runs on a worker; normal quit flushes queued writes, while forced process termination can lose uncommitted writes.
Version 0.10 offers an explicit fast setup button to enable Live upload, 20-second connection reuse and automatic field insertion. History remains off. Sign-in launch is an explicit installer/in-app choice stored in the current user's Windows Run registry entry; it never starts microphone recording by itself. Setup requires an audio-free connection check and an explicit local microphone test. Uninstall preserves the user's settings, key and opt-in history.
Updated for 0.11.4, 8 November 2026: Remove saved key cancels current operations, clears retained provider connections/checks and removes only the current DPAPI credential. It leaves legacy imports, settings and history unchanged and does not revoke the key at Google. The free preview has no ASWorks activation or payment requirement. ASWorks download servers process ordinary download requests; the app recognition requests go directly to Google. Public privacy and support: https://asworks.studio/apps/veylo/privacy and https://asworks.studio/apps/veylo/help. Contact hello@asworks.studio.