Your voice stays yours.

The microphone stays closed while Saral is idle. It opens when you start dictation or a meeting recording, and closes before transcription begins, after you cancel, after an error and when you quit.

• There is no background pre-buffer. Nothing is captured before you press your shortcut.
• Model discovery and update checks never open the microphone.
• Live dictation is limited to ten minutes of buffered audio.

On-device. With a downloaded Whisper model, recognition runs on your computer and your audio isn’t sent anywhere. The Lightweight local preset also removes cloud fallback and AI routing.

Cloud providers. If you connect a provider such as Groq, Deepgram, ElevenLabs or Sarvam, your recording is sent to that provider with your key, under their terms. If you use AI polish, the transcript text goes to the writing provider you picked. Saral doesn’t route requests through its own servers.

• Automatic fallback only tries providers you’ve configured, in the order you set.
• The experimental “Transcribe during pauses” option is off by default and can make extra cloud requests.
• Meeting recordings are sent in short sections to cloud speech providers; live meetings skip silence on-device first.

Keys are saved in macOS Keychain or Windows Credential Manager. Saral’s settings file only holds references to them, and keys are masked in the interface.

• If the credential store isn’t available, Saral keeps your existing file and never falls back to writing keys in plain text.
• The window talks to Saral’s engine through a private service bound to 127.0.0.1 on a random port, with a new token each launch.
• Feedback emails are drafted for you to review and send. No keys, audio, transcripts or logs are attached.

Dictation history keeps your latest 50 transcripts in memory for the current session. If you turn on local storage, up to 500 transcripts are kept for 1, 7, 30 or 90 days in ~/.flowclone/history.json with owner-only permissions. Turning it off removes that file.

Meetings keep transcripts and notes locally (up to 100 meetings and 12 MB of text). Temporary audio is deleted after processing, and files you import are never touched.

Learned vocabulary holds at most 128 terms plus aggregate timings, in a file capped at 64 KB. It never contains recordings or full transcripts, and Reset learning removes it.

• A little of the text around your cursor can be read to get spacing and capitals right. It stays on your device.
• Saral doesn’t capture your screen.

Release builds share anonymous, aggregate counts so we can see whether Saral is working well: things like how many dictations and meetings started or finished, seconds of audio processed, processing time, word counts and broad app families. Unknown apps are counted as “other”.

An installation is identified by a random ID, which the server stores only as a hash. It counts installations, not people.

• Never collected: your IP address (in the analytics database), hardware IDs, accounts, API keys, audio, transcripts, clipboard contents, URLs, window titles or raw app names.
• Turn it off in Settings → Privacy & data → Share anonymous usage counts. Unsent totals are cleared straight away.

• The cloud providers you configure, using your keys.
• Update checks against public GitHub release data and updates.saralai.xyz. Installers are verified before you open them.
• Model catalogs from connected providers, refreshed in the background. Discovery never changes your choices or downloads weights.
• Anonymous usage counts, if you leave them on.

Questions or concerns? Write to support@saralai.xyz

This page describes how the app behaves. It isn’t a legal privacy policy. Last updated 10 October 2026.