Direct answer
On-device speech translation can reduce the need to send speech content to a remote inference service, but it does not automatically make the complete app private. Microphone permission controls capture, while model execution, audio buffers, transcript history, crash reports, operating-system backup, clipboard, synthesized playback, and user-selected sharing are separate data boundaries. A privacy assessment should trace each of those stages from a user action to storage or network destinations. Cove's current inference fact card describes a local audio path through LiteRT-LM; that implementation fact does not establish that every optional service or future code path makes no network request. For a trustworthy “on-device” statement, inventory the data, inspect traffic and logs, test backup and deletion behavior, and distinguish user-requested export from automatic transfer. The strongest useful claim is narrow: which content stays local during which tested task, which setup connections happen before it, and what leaves the device only after an explicit user action.
Scenario: translating a private voice request in a taxi
A traveler records a sentence that includes a hotel name and a personal appointment time. The app translates it locally, saves a history item, and offers a speak button. The traveler later copies the result to a messaging app and has system backup enabled. Calling the model local answers only one part of the privacy question.
Decision
Trace the source recording and transcript through capture, inference, history, diagnostics, backup, playback, clipboard, and share. Identify which actions are automatic, which require consent, and which destination is outside the translation app.
Expected outcome
The privacy notice can separate what remains on-device during inference from setup downloads and user-directed sharing. The user can delete the original and derived data without assuming that hiding a history row removed backups, logs, or exported copies.
Evidence to collect
The claim becomes reviewable only when the following evidence is attached to the same artifact and test run:
- Audio and transcript data inventory
- Per-stage traffic and log review
- Export, backup, and deletion verification
A privacy claim needs a data inventory and traffic evidence. A local-model label by itself is not proof.
Implementation workflow
- Build a data inventory for every speech stage. List microphone samples, encoded audio, recognized text, translated text, language selections, timestamps, model identifiers, file paths, and diagnostic fields. For each item record whether it exists only in memory, is written to app storage, appears in history, enters a log, or is included in an export. Include derived transcripts and cached audio, not just the original recording.
- Map user action to code path and destination. Follow a voice task from permission request through capture, model call, result parsing, history persistence, playback, copy, and share. Enumerate every SDK or endpoint contacted during setup, inference, crash reporting, purchases, and newsletter features. “The model runs locally” describes inference location; it is not proof that no other request contains content or identifying metadata.
- Inspect network behavior without private test data. Use synthetic, non-sensitive recordings and capture traffic during first launch, model download, a translation, a failed task, diagnostics, and export. Compare those traces with the declared inventory. If the test fails because connectivity is unavailable, record whether the task itself failed or only an optional service failed. Do not put real personal speech into packet logs or public test artifacts.
- Test logs, crash paths, and operating-system backup. Trigger parse errors, cancellation, storage failure, and app recreation with synthetic samples. Confirm that logs contain stage codes and safe identifiers rather than audio bytes, transcript text, or private file paths. Inspect backup configuration and perform a restore test. A local database can still be backed up by platform services depending on app configuration and user settings.
- Treat playback, clipboard, and sharing as separate choices. Check which TTS engine reads the translated text and whether it is available offline; do not infer its network behavior from the local model. Verify that copy and share happen only after a clear user action and show the exact content being handed to another app. Explain that after sharing, the destination app and its sync policy control the data.
- Verify deletion across original and derived data. Create a test record and delete it through the user interface. Check the source audio, transcript, translated output, temporary buffers, caches, search indexes, logs, backup, and exported file separately. Document which copies the app can remove and which require the user to act in an operating-system backup or third-party destination.
Keep each transition observable. A failure should identify the stage, artifact, runtime, and recovery action without logging private user content.
Acceptance criteria
- Each audio, transcript, translation, metadata, and derived artifact has a documented location and retention behavior.
- Traffic and log inspection cover setup, inference, failure, diagnostics, backup, and export paths using synthetic data.
- Local inference, optional TTS, operating-system backup, clipboard, and sharing are described as separate boundaries.
- Deletion tests verify original and derived artifacts rather than only hiding a history item.
Decision table
| Dimension | Choice or evidence | How to interpret it |
|---|---|---|
| Inference | Audio and prompts are processed by a local model after setup | Reduces remote inference transfer for the tested path; does not establish zero network traffic. |
| Setup and diagnostics | Model downloads, crash reports, or account services are audited independently | These can contact remote systems even when a translation runs locally. |
| User sharing | Copy, export, and share require visible user action | The destination app or storage provider becomes a separate privacy boundary. |
| Retention | Original audio, transcript, translation, cache, and backup are tested separately | Deleting one screen entry does not prove every copy was removed. |
Failure patterns to prevent
- Treating local inference as zero data flow
- Logging raw transcripts or file paths
- Calling optional TTS or cloud backup private without checking
Also prevent silent fallback, unpinned artifacts, missing cancellation, and conclusions that combine unlike configurations. Store unsuccessful runs alongside successful ones.
Minimum reproducibility record
| Layer | Record |
|---|---|
| Device | Manufacturer, model, chipset, RAM class, operating-system build |
| Software | Application version and git commit |
| Model | Family, variant, revision, format, file length, hash, quantization |
| Runtime | Name, revision, requested backend, observed backend evidence |
| Workload | Fixture revision, input hash, prompt hash, output policy |
| Outcome | Completed, failed, cancelled, fallback, and privacy-safe diagnostics |
Release checklist
- ☐ The primary query is answered without an unsupported number.
- ☐ Every artifact and runtime is pinned.
- ☐ The representative task and failure policy are explicit.
- ☐ Lifecycle, cancellation, cleanup, and fallback are tested.
- ☐ User-content and network boundaries are documented.
- ☐ Result wording applies only to the recorded configuration.
- ☐ The page links to raw method or evidence when results are added.
Sources and related evidence
- Cove Travel voice translation
- Android privacy and security
- Android guidance on requesting microphone access
- Android privacy and security overview
- Android guidance on sharing microphone input
Chinese deployment and troubleshooting content is organized in the .奇连 AI 端侧专题.
Frequently asked questions
What privacy benefit does on-device speech translation provide?
When audio inference truly runs on the device, the app may not need to send that speech to a remote inference server for the tested task. That can reduce one important exposure, but the claim must be checked against the actual code path and traffic. It does not automatically cover model downloads, analytics, backups, logs, TTS, or sharing.
Does local inference mean the app makes no network requests?
No. Installation, model delivery, optional analytics, purchases, newsletter signup, crash reports, and user-directed sharing may use other network paths. Inventory each request, its trigger, and fields. Test the core task with connectivity disabled only after required setup, then inspect reconnect behavior separately.
Can a speech transcript be more sensitive than the recording?
Yes. A transcript may expose names, plans, addresses, or appointment details in a searchable database or crash log. Treat audio, transcript, translation, metadata, and derived indexes as distinct content-bearing data and apply appropriate access, retention, redaction, and deletion rules to each.
What should I check before sharing a translated sentence?
Review the exact text, trigger sharing explicitly, and identify the app or destination receiving it. Explain that once the content enters another app, that app may store or synchronize it under its own settings. Local model execution does not make a later user-directed export stay inside the original app.