Offline Speech-to-Text with Whisper

MyBenAI transcribes speech with OpenAI's Whisper, running locally through whisper.cpp. The audio is turned into text on the phone. It is never uploaded, there is no per-minute quota, and there is no monthly bill for a transcription service.

Which Whisper you get

The device profiler picks the largest that fits: tiny on a floor-tier phone, base on 4 GB, small on 8 GB, and large-v3-turbo on a flagship. All are quantised and multilingual, and tiny is only 30 MB — small enough that even the cheapest phone in the catalog gets working dictation. You can install a different one from the model manager if you would rather trade speed for accuracy.

What it is good for

Dictating a message rather than typing it. Transcribing a voice memo. Turning a recorded meeting or lecture into text you can then ask questions about. That last one is the case where local matters most: a recording of a meeting is exactly the kind of file people are contractually not allowed to hand to a transcription service, and here there is no service.

Beyond dictation

Whisper transcribes far more languages than the interface is translated into, so speaking a language the app does not display still works. Pair it with hands-free mode to hold a conversation without touching the phone, and with read aloud to hear the answer.

Memory behaviour

The Whisper context is released after sixty seconds idle, rather than sitting in RAM between utterances — the same discipline that applies to every heavy model here. You will not notice; the phone will.