Murmurtale The default: nothing leaves your Mac
On the free tier, Murmurtale runs entirely on your device. After a one-time model download, your microphone audio is transcribed and cleaned up locally — by an on-device speech model and on-device formatting. No audio and no text are sent to us or to anyone else. You can dictate up to 10,000 words with no account at all, and a free account then gives you 10,000 local words every month, at no cost — the account only ever sees a monthly word count, never your dictation.
That includes the parts that feel like they should need a server. Suggesting new dictionary terms and learning from your corrections both happen on-device, over your local data. It is not keystroke logging, and nothing is uploaded to produce those suggestions.
What lives on your device
A few things are stored locally, under your control, and never uploaded to Murmurtale’s servers:
| Data | Where | How long |
|---|---|---|
| Dictionaries (your names, places, terms) | Your Mac | Until you delete them |
| History of dictations | Your Mac | You choose: off, 1, 7, or 30 days (default 30); clear anytime |
| Recent audio buffers (for retry and playback) | Your Mac | Auto-pruned with History |
| Settings (genre, cleanup level, hotkeys) | Your Mac | Until you change them |
If you use Apple’s iCloud Drive with Desktop & Documents sync, macOS may sync these files within your own iCloud account under Apple’s terms — that is your Apple account, not Murmurtale’s servers, and you can turn it off in System Settings.
When the cloud engine helps — and what it does
Some audio is genuinely hard: heavy background noise, a tricky accent for the local model, a passage the on-device pass isn’t confident about. For those cases, Pro offers an optional cloud engine that can improve transcription and dialogue handling. It is opt-in — never on unless you turn it on — and if a cloud request ever fails, Murmurtale falls back to the on-device engine automatically so you are not left waiting.
When you do use it, here is the honest data path:
- Your audio and/or transcript are sent over an encrypted connection (TLS 1.3) to the Murmurtale backend, which forwards them to processing subprocessors.
- On the Murmurtale backend, content is processed in memory only and is never logged or persisted; we store no dictation content on our servers and do not use it for analytics or model training. Our cloud subprocessors process it transiently too — the details are in the next section.
- The backend keeps only minimal non-content operational metadata — timestamps, request and byte counts, error codes, your account ID and tier — to run and bill the service. That metadata does not contain your dictation audio or text.
What our cloud providers may keep — and what they never do
Our rule for any cloud provider is the same and non-negotiable: they may process your content only transiently to return the result, may retain it only briefly — up to 30 days — for trust-and-safety review before automatic deletion, and may never use it to train models. Here is the current roster.
| Subprocessor | Purpose | Handling |
|---|---|---|
| Anthropic (Claude) | Cloud cleanup / editing pass | Processed transiently; may be retained up to 30 days for trust-and-safety before automatic deletion; never used for training |
| AssemblyAI | Cloud speech-to-text | Murmurtale deletes each transcript the moment your result is returned; audio processed transiently; never used for training |
The current authoritative version of this list, with any changes, lives in the Privacy Policy.
Diagnostics, and what they never include
Murmurtale can send privacy-safe diagnostics for crashes and handled failures — app version, OS and device class, account tier, which pipeline stage failed, sanitized error codes, and small crash fingerprints. Diagnostics never include your dictated text, audio, prompts, dictionary terms, manuscript content, email, or account secrets. You can turn them off in Preferences.
Why build it this way
Because the honest default for a writing tool is that your unfinished work stays yours. Local-first means the privacy story is simple: we cannot leak content we never receive. The cloud engine exists for the moments you want more accuracy, on your terms, with retention set to zero.
Read more
- Local vs cloud — choosing an engine in the app.
- Privacy and authorship — the in-app promise.
- Privacy Policy — the full, versioned document.