Diktafon is a voice-memo app: each topic is a cassette, memos are appended
to its tape, and a cassette plays back as one continuous recording. Every
memo is transcribed and every cassette summarised, fully on-device:
whisper.cpp for transcription, Qwen3 via llama.cpp for summaries — vendored
behind two small C shims sharing one ggml build, driven from Dart (Flutter)
over FFI in worker isolates. A bundled Silero VAD keeps silence and noise
from turning into invented sentences. Models are downloaded once,
sha256-pinned. No accounts, no cloud, no analytics. MIT; Android, iOS,
Linux. It came from a real need. The closest existing apps are WalkWrite (iOS-only, aimed at writers — speech becomes polished text) and Notely Voice (a notes app fed by dictation); both turn audio into text, while I wanted the recordings to stay the primary object, with the text attached. Although I'm an experienced programmer, I wrote almost none of this by hand. I kept to the role of designer and reviewer, and Claude (Fable) wrote the code — an experiment in how far that gets. It did as I expected — very well, app complete in about a week of occasional 30 minutes sessions. Happy to answer questions about that workflow as much as the app. I use it daily and I'm happy with it. I'll be glad for any feedback. |