Privacy

Wavelet transcribes on your Mac with an on-device model. Your audio stays on your Mac.

While you record, Wavelet writes the raw audio to a file on your Mac, even when transcript history and the audio archive are off, so a crash or a cancel does not lose the recording. When the recording is transcribed, the file is deleted. If you cancel, if transcription fails, or if Wavelet stops mid-recording, the audio is kept for 30 days by default, so you can transcribe it later from History. You can delete it sooner in History or Settings.

The 30 days count from the cancel or the failure, or, after a crash, from the next time Wavelet starts. A failed recording kept by an earlier version gets its 30 days from the first launch of the update. Once they are up, the audio is deleted the next time Wavelet starts, shows History, or finishes a recording, unless Wavelet is recording or recovering audio at that moment; then it waits for the next of these. The files hold the audio and the recording's technical details (such as the microphone name, sample rate, and timestamps), never text. An hour of recording takes about 0.7 GB (1.4 GB in stereo). Clear all history in Settings removes them. Nothing in them is sent to Littoral AI.

Transcript history stays on your Mac. During setup, you choose whether to keep it, and you can clear it at any time. Wavelet also keeps recent recordings for 30 days by default so you can transcribe a clip again; in Settings you can turn this off or keep them until you clear them. Littoral AI does not collect your recordings or transcripts. If you turn on Dreamer, your transcripts are relayed through Littoral AI's AWS server in Canada to Anthropic, as described below.

Dreamer is an optional helper. It looks back over your transcripts, finds the words Wavelet kept getting wrong, and proposes them for your dictionary, so Wavelet gets better over time without you adding words by hand. It is off until you turn it on.

Once a day, Wavelet sends the transcripts still in your History, via a relay AWS server in Canada, to Anthropic in the United States, where the Claude model reads them. It also sends your dictionary words, replacement rules and snippet triggers, so Claude can see which rules were already applied to each transcript and what is still missing. Anthropic's commercial API terms do not allow it to train its models on them. Wavelet never sends recordings. You can turn Dreamer off at any time.

You can use Wavelet without an account, sign-up, or email address.

If you buy Wavelet, Stripe handles your payment and checkout email. Littoral AI receives the checkout email, order record, and payment status to deliver, recover, or refund your license.

Purchase records are kept only as long as needed for license delivery and recovery, refunds, accounting, and legal obligations. They are never joined to the random ID used for usage statistics. For a privacy request, email wavelet@littoralai.com.

Beta and trial builds send minimal, anonymous usage statistics that carry no content. They are on by default. You can turn them off at any time in Settings > Usage statistics.

They include:

• A random install ID, the app and macOS versions, and the broad type of Mac.
• Anonymous information, with no content, about feature use, reliability, and performance: daily counts of dictations and their outcomes, timing and memory readings for each dictation, how many days since installation, whether the macOS permissions Wavelet relies on are granted, and which days the app was running, at most once per day.
• When a dictation fails, a short code for the kind of failure, drawn from a fixed list, never an error message or log text.
• The engine, outcome, duration, cost and number of suggestions from Dreamer runs, and daily counts of suggestions kept or dismissed; in beta builds, each model's calibration word error rate; and, when statistics are enabled at purchase, that a license was activated and on which day of the trial.

They never include: audio, recordings, transcripts, clipboard contents, dictated words, suggestion text, app or window names, process or app lists, raw logs, location, IP address, a license or license key, email, order details, payment information, or anything that identifies you.

Statistics go to PostHog on its EU servers with IP capture turned off. Littoral AI keeps aggregate counts only while they are useful for improving Wavelet.

Activating a license does not change your saved setting or random install ID. If statistics were off, they stay off. If they were on, they stay on. Delete my usage data replaces the random ID on your Mac; it cannot erase anonymous statistics already sent to PostHog.

Our servers keep standard access logs, as any website does, which we use for aggregate counts by region.

Last updated October 2026, during the Wavelet beta.