Building a Privacy-First iOS Voice Keyboard Without Background Analytics

Every Voice Keyboard I Tried Was Running More Than Transcription. So I Built One That Doesn’t.

Developer Ondrej Machala created Diction to address privacy concerns in mobile dictation software. Most voice keyboards bundle analytics SDKs and attribution libraries that run in the background during every keyboard session.

Why This Matters

In the ideal model of a transcription service, the system only processes audio when requested. However, technical reality reveals that many apps use crash reports to capture text fragments from memory and analytics to track user metadata. By stripping these non-essential services, Diction ensures that voice data—which is often more candid and personal than typed text—remains under the user’s control rather than being harvested by third-party SDKs.

Key Insights

• Crash reports in standard voice keyboards can inadvertently include fragments of text from device memory (Ondrej Machala, 2026).
• On-device transcription mode allows for zero-network requests, ensuring audio never leaves the iPhone.
• Self-hosted transcription pipelines enable users to control the server code, which is hosted on GitHub for public audit.
• Diction One cloud processing handles audio in real-time without storing the processed data on the provider’s side.

Practical Applications

• Use Case: Deploying a self-hosted transcription server to maintain total data sovereignty for corporate environments. Pitfall: Relying on third-party attribution libraries can lead to metadata leakage.
• Use Case: Using on-device mode for dictating sensitive personal notes without any network connectivity. Pitfall: Capturing metadata via analytics SDKs can compromise the candid nature of voice dictation.

References:

• https://dev.to/omachala/every-voice-keyboard-i-tried-was-running-more-than-transcription-so-i-built-one-that-doesnt-e49