What it does

Record, transcribe, replay, keep what you choose

This page describes how the app works today in private testing. “Current” features are in the app now and may change before release. “Planned” features are not built yet and have no dates.

1. Recording

Current

  • Recording runs as an Android foreground service with a visible notification while the microphone is in use.
  • You start it by tapping Start. After a restart of the phone, a force-stop or an interruption, recording does not resume by itself: it needs your tap.
  • Audio is saved in pieces of about 30 seconds, written into an encrypted journal on the phone, never as plain audio files. The app records as compressed audio (AAC, the default) or uncompressed (WAV), which takes much more space. It is kept out of Android's cloud backup.
  • Recording continues while earlier audio is still being transcribed.

2. Transcription on the phone

Current

  • Saved pieces are queued and turned into text by a speech model running on the phone. There is no cloud transcription. A piece that fails is retried a limited number of times and then marked as needing attention. It is not silently dropped.
  • It is not instant, and not live captions. Audio is handled in 30-second pieces with a short look-ahead, so text cannot appear sooner than roughly half a minute after you speak. How long it really takes depends on the phone and on what else it is doing.
  • Recognition makes mistakes, more so in noisy rooms and with names and numbers. You can correct the text; the original stays in its history.
  • No model is bundled in the app. Models are downloaded once from Hugging Face, on Wi-Fi by default, and checked against a published hash, or you can import one from a file. See the privacy page.

3. Conversation playback

Current

Each piece of text stays linked to the audio it came from, as long as that audio is kept.

  • Play a whole recording, or just the part behind one transcript entry, along a timeline you can seek with your finger, with pause, resume and stop.
  • Each piece of audio sits at the time it was captured, so the length shown is the recorded time.
  • Play a conversation made of several recordings. The time between recordings is left out and marked “Between recordings”, instead of playing silence.
  • When audio is not available, the player says why: expired, missing, still being written, or never recorded.
Illustration of a playback timeline A bar made of a recording, a short missing section, a second part of the recording, a marker for the time between recordings, and a second recording. unavailable Recording 1 Recording 2 Between recordings: time omitted
Illustration of the timeline, not a screenshot.

4. How long audio and text are kept

Current

Audio age limit
By default audio is kept for a rolling 4 hours. You can choose 1, 2, 4, 8 or 24 hours, or “Keep until I delete”, which turns off age-based expiry.
Storage limits still apply
A separate audio capacity (1 GiB by default) and a free-space reserve of the phone can still pause recording, even with “Keep until I delete”. The app shows how much of the capacity is used, warns from about 80%, and stops new recording at 100% to protect the phone. It says why when it pauses.
Not yet transcribed
Audio waiting to be transcribed gets a separate recovery allowance so that it is not lost just because it is old. If that allowance fills, recording pauses.
Pinned audio
You can pin recent audio so that it is kept.
Transcripts
Transcripts stay in your encrypted library until you delete them, even after their audio has expired.
Deleting
Deleting a transcript entry shows how many audio pieces and neighbouring statements it affects before you confirm.

5. Correct, search and share

Current

  • Edit transcript text. The original machine text and the edit history are kept, and you can undo or restore the original.
  • Search your transcripts, including corrected text.
  • Copy, share or save text, audio or a ZIP of what you select. Cancelling the share sheet sends nothing.
  • Make your own encrypted backup file and restore from it.

Planned

Planned These are not built yet and have no dates.

  • One-time purchase. Saidwise is planned as a one-time purchase, with no subscription. The price has not been announced.
  • More languages and translation. Dutch recognition and English to Dutch translation on the phone, and further languages later. Machine translation can be wrong and will be labelled as machine-made.
  • Speaker names. Telling speakers apart, naming them and remembering them.
  • Summaries. Short summaries made on the phone that point back to the passages they come from.
  • Ask Saidwise. Typed or spoken questions about a conversation, answered with the passage they came from, and saying so when the transcript has no answer.
  • Highlights. Mark a passage as important for a summary, or just for your own reading.

Good to know

  • Saidwise needs Android 12 or newer on a 64-bit ARM phone. There is no iPhone version.
  • We have not published accuracy or speed figures, and we do not promise that it works on every phone.
  • Recording and on-device transcription use the processor and battery. We have not published battery figures.
  • Audio is kept for the time you choose, within a storage limit. Saidwise is not a place to keep recordings forever.