GrapheneOS launches Speech Services v2 with privacy‑first enhancements

GrapheneOS Speech Services → Better on‑device voice handling

The GrapheneOS team announced the release of Speech Services v2 on their GitHub repository (release page). The update is a direct response to user demand for voice‑driven interactions that do not rely on cloud endpoints, a core tenet of GrapheneOS’s privacy‑first philosophy.

What the new version does

• On‑device language models – The most visible change is the inclusion of small, locally stored acoustic and language models. These enable basic speech‑to‑text conversion without sending audio to external servers. The models are deliberately lightweight (under 30 MB) to keep the OS footprint modest while still supporting the most common languages.
• Granular permission gating – Apps now must request the android.permission.RECORD_AUDIO permission and explicitly opt‑in to the on‑device Speech Services API. This double‑check prevents silent activation by background processes.
• Improved API stability – The public interface has been cleaned up to remove deprecated callbacks and to provide clearer error codes. Developers can now query the service for model availability, which helps them fall back to alternative input methods when needed.
• Battery‑aware processing – The service monitors device power state and throttles inference when the battery is low, reducing the risk of unexpected drain during prolonged voice sessions.

Why it matters for the ecosystem

GrapheneOS has long positioned itself as a security‑focused alternative to mainstream Android builds. By moving speech recognition entirely onto the device, the project eliminates a common privacy blind spot: the transmission of raw audio to third‑party providers. This step aligns with the broader trend of edge AI, where inference runs locally to avoid latency and data leakage.

For developers, the updated API means they can build voice‑enabled features without compromising the privacy guarantees that GrapheneOS users expect. The clear permission model also simplifies compliance reviews, as there is less ambiguity about when audio data might leave the handset.

Community reaction and next steps

Early feedback on the GrapheneOS discussion forum highlights appreciation for the offline capability, especially among users of privacy‑sensitive devices. Some developers have already begun integrating the new API into custom keyboards and accessibility tools.

The release notes point to a full changelog that details bug fixes and performance tweaks. The team has indicated that future iterations will expand language coverage and explore on‑device natural‑language understanding, which could eventually support voice commands beyond simple transcription.

---

For a quick look at the changes, see the official GitHub release notes.