Blog
Why Talki saves your recording before it transcribes
A dictation tool usually treats audio as a means to an end. We think the audio is the thought, and the text is one attempt at writing it down.
The moment that started it
You have an idea while walking to the kettle. You press a shortcut, say three sentences, and the text that comes back reads like someone else's notes. A name is wrong, a clause is missing, a number has changed. The audio is gone, and so is half the idea.
Speech recognition has become very good, but not perfect, and the mistakes are rarely random. They land on the words that matter most to you: names, product terms, the unusual word you chose on purpose. When the recording is discarded, the error becomes the record.
The rule
Talki follows one ordering rule:
- Stop recording.
- Save the original audio to the library on your Mac.
- Only then transcribe.
Everything else in the app builds on that. A network error is a retry, not a loss. A wrong model is a retry with a better one. A transcript you doubt can be checked against what you said.
Three layers, kept apart
Each capture in Talki keeps three things side by side:
- The original recording, exactly as captured.
- The original transcript, exactly as the provider returned it, with the provider and model noted.
- The edited text, after your dictionary, optional cleanup and your own edits.
Editing the third never changes the first two. That makes cleanup safe to try: if an AI rewrite drifts from what you meant, the version you actually said is one click away.
What it costs
Keeping audio takes disk space. Short dictations are small, and Talki stores them in a local library you can prune: moving a note to Trash is reversible, and deleting it for good asks first.
It also costs a little speed in the interface, because Talki waits for the file to be written before it reports "saved". We think a word that says saved should mean the file exists.
What we chose not to do
- No silent cloud backup. Recordings stay on the Mac that captured them. The iCloud sync we are building carries text, not audio.
- No automatic paid retries. If a cloud request fails, you decide whether to spend another request.
- No meeting recording. Talki is a tool for your own voice, not for capturing other people.
Where this goes next
Keeping the source opens doors we want to walk through carefully: re-transcribing old notes with better models, comparing two transcripts of the same recording, and moving ideas between devices without losing where they came from. Each of these depends on one thing being true: the original is still there.
If you have a story about a dictation that went wrong, we would like to hear it at support@talki.im.