Privacy policy.

The short version

Silo Systems collects no data from OhmsVoice and operates no OhmsVoice server.

No analytics. No telemetry. No account. No usage metrics. No error beacons. No advertising IDs. No fingerprinting. No crash reports. No app data is sent to Silo Systems.

Recording and transcription run on the Apple device where you use the app. Raw Transcript and Clean (On-Device) never leave the device, and snippets expand locally. If you save an API key and explicitly allow that provider, a cloud rewrite preset sends the transcript text, the preset's rewrite instructions and any spellings you saved, never the audio recording, directly to that provider under your account. Through version 1.1.2, a rewrite started on a Mac also includes the name of the app you were dictating into. On a Mac, a local model server you choose is reached only on loopback. Nothing is proxied through or received by Silo Systems.

What we don't have, and therefore never see

Because OhmsVoice has no server-side component, none of the following exists anywhere in our infrastructure (which consists of: nothing): your name, email, phone, or address; your IP address, device identifier, or any hardware fingerprint; your content and anything derived from it; session timing, feature usage, or behavioural data; crash reports. The app does not call home to Silo Systems. A model download, provider rewrite, App Store update, or link you open can contact the relevant third party exactly as described below.

Outbound traffic you initiate

These are the only network paths. They are user-initiated, and none route through Silo Systems:

1. Model weight download. If you choose to download the open speech model, in the first-run setup card or in Settings, OhmsVoice fetches the pinned Parakeet files (about 483 MB) from Hugging Face and its download CDN, then verifies every file against a bundled SHA-256 checksum and load-tests the model before marking it ready. No Ohms account or token is sent. As with any HTTPS request, the host can receive ordinary connection metadata such as your IP address under its own privacy policy. Once installed, transcription with that model runs offline. If you choose built-in recognition instead, nothing is downloaded.
2. Optional cloud rewrites. Cloud presets remain blocked until you save a key and approve the named provider in Settings. When you then process a dictation with that preset, OhmsVoice sends one request directly to the provider you made active: OpenAI, Anthropic, OpenRouter, Google Gemini, or xAI (GitHub Models too, through version 1.1.2). It carries your API key, the transcript text, the preset's rewrite instructions, and any spellings you saved in Settings → Dictation → Spellings. Through version 1.1.2, a rewrite started on a Mac also includes the name of the app that was in front when you started dictating, plus tone and formatting hints derived from it; later versions never send it. The audio file is never sent. The provider also receives ordinary request metadata such as your IP address, under your account and its privacy and retention terms. Deleting that key in Settings also revokes OhmsVoice's stored consent for that provider.
3. Local model server (Mac only). If you select Local Ollama on a Mac, OhmsVoice sends the same text only to 127.0.0.1:11434 on that Mac. That loopback request does not go to an external service. iPhone and iPad do not offer this option.
4. App Store updates. When Apple ships an OhmsVoice update, the App Store delivers it. This traffic is between your device and Apple's servers. OhmsVoice is not involved.
5. Links you click. If you click an external link inside the app, your default browser handles the request. OhmsVoice is not involved.

Version 1.1.2 and earlier. In those versions, a cloud rewrite also sent the name of the app you were dictating into, a short tone and formatting hint derived from that app, and a built-in list of product-name spellings, and GitHub Models was offered as a provider. Releases after 1.1.2 remove all of these, and the paths above describe them.

Permissions we request, and why

• Microphone. Required only while you record a dictation. Audio is transcribed locally and never sent to a rewrite provider.
• Speech Recognition. Requested only if you choose built-in recognition. OhmsVoice requires on-device recognition and never sends audio to Apple; if on-device recognition is not available for your language, it tells you instead.

The Mac shortcut is a standard system hot-key registration and needs no Accessibility or Input Monitoring permission. OhmsVoice does not request protected resources beyond those listed above. System permission controls remain in effect at all times.

On-device data you create

Audio is held in a temporary file while a dictation is processed and is deleted after processing or cancellation; crash leftovers are swept at the next launch. Optional transcript history follows the retention period you choose and can be cleared in the app. The name of the app you dictate into is read on the device for per-app preset rules and History; version 1.1.2 and earlier also sent it with cloud rewrites, as noted above. Provider keys stay in the system Keychain; consent choices, snippets, saved spellings, per-app rules and settings stay in local preferences; and the downloaded speech model stays in the app's sandbox. On iPhone or iPad, deleting OhmsVoice rather than offloading it removes app-sandbox data. On Mac, use the in-app delete and purge controls before removing the app if you also want its sandbox data removed. Silo Systems cannot access any of it.

Third-party components

Local components receive no data over the network. Optional services are contacted only through the paths described above:

• FluidAudio (vendored, Apache-2.0): runs locally on-device.
• Parakeet TDT 0.6B v3 CoreML (CC-BY-4.0): the open speech model, runs locally on-device.
• Apple's built-in speech recognition (on-device, optional): runs locally on-device.
• Hugging Face and its download CDN: contacted only when you choose to download the pinned open speech model.
• OpenAI, Anthropic, OpenRouter, Google Gemini, or xAI: contacted only after you save a key, explicitly allow that provider, and run a cloud preset. Their own privacy and retention terms apply.
• Ollama (Mac only): optional local service contacted only on loopback at 127.0.0.1:11434.

Children's privacy

OhmsVoice is rated 4+, and Silo Systems does not knowingly collect data from users of any age. Optional cloud providers have their own eligibility, privacy, and account terms. A parent or guardian should not enable one for a child unless that use complies with the provider's terms.

Changes

If we ever materially change how OhmsVoice handles data, we will publish an updated version of this page, announce the change prominently in the release notes, and require explicit consent in-app before any new transmission begins. The current policy is: Silo Systems receives nothing, and external traffic is limited to the user-controlled paths above.

Contact

Questions about this policy: admin@ohmslaw.net