A new offline-first AI dictation app from Google has appeared on iOS, built around the company’s Gemma models and positioned as a direct challenge to tools like Wispr Flow. The app records speech, runs transcription locally, and then syncs results when a connection is available, keeping the core language processing on-device.
Gemma, a family of compact large language models derived from transformer architectures and optimized for low-latency inference, allows the app to decode audio without routing every request to cloud servers. That design reduces network dependence and enhances data minimization, a key concern in voice interfaces used for email drafts, notes, and productivity workflows on mobile devices.
By focusing on offline performance, Google targets scenarios where connectivity is unreliable or where users want stricter control over their acoustic data. The release also strengthens Google’s position in the broader speech recognition and natural language processing stack on Apple’s platform, in direct competition with third-party dictation apps that still lean heavily on remote compute.
