Ciji Transcribe: Voice to Text

Chang Yuan-Yi

Contains ads

About this app

You already have the recording. This turns it into text.

Bring in a meeting, a lecture, an interview — a file from a phone, a watch, a voice recorder, or one somebody sent you — and read it back as words with timestamps. The recognition runs on the device you are holding. Nothing is uploaded, and there is no account.

---

Read it, and find the moment

• Every line carries the time it was said; tap one and the audio plays from there
• Playback from half speed to two and a half times, for the parts worth slowing down
• Skip back or forward ten seconds without losing your place
• Select and copy any line straight out of the transcript

Who said which line

• Work out how many people were talking and attribute each line to one of them
• Runs as a separate step on a transcript that already exists, so nobody waits for it who does not want it
• No length limit — an hour-long meeting costs more time, not more memory
• Two ways to read the same transcript: plain, or with every line marked

Fix what it got wrong

• One person split into several? Tick them and merge them back into one
• Give a speaker a name and it is used everywhere, including in exported files
• Move a single line to somebody else when a short interjection landed with the wrong voice
• Start the speaker step over without touching a word of the text

Take it with you

• Export as plain text, timestamped, one line per segment
• Or as a subtitle file that plays alongside the original recording
• With the speakers marked, or without — a transcript is read for what was said, and who said it is a second question
• Goes out through the sharing sheet, so it can be saved, mailed or sent anywhere

Choose how it listens

• Eleven recognition models to download, from small and quick to large and careful
• Each one says what languages it handles and what it does poorly, so the choice is yours to make
• Run a second model over the same recording and keep both results side by side
• Delete a model when you want the space back; download it again whenever

Long recordings, handled

• Pause a transcription and pick it up later from where it stopped
• Leave the screen and it keeps going
• Progress is shown as position in the audio, with an estimate of what is left
• The screen stays awake while it works, because a locked phone is a stopped job

---

What runs where

• Recognition and speaker identification run on your device, in your hands
• Your recordings are never uploaded, and there is no account to create
• Models are downloaded once, from an open-source project, and then work without a network
• The app is free and supported by ads
• Every model and library it builds on is credited inside the app, with its licence

Worth knowing before you download

• The app does not record audio. It reads recordings you already have.
• A recognition model has to be downloaded before the first transcription — a few hundred megabytes, kept until you delete it.
• Chinese output is normalised to Traditional characters by a table built into the app.

Good for:
• Reading back a meeting instead of listening to it again
• Finding the one thing that was said, by its timestamp
• Handing somebody a written record of a conversation
• Subtitling a recording you already made

What’s new

- Smoother, more reliable experience with under-the-hood improvements.