Ask a question across your recordings
Somebody agreed to something three weeks ago and nobody wrote it down. This transcribes the recordings, answers a question across them, and tells you where to listen.
- Answers carry a timecode you can play from, not just a quotation.
- A timecode cannot be invented, because the model is never told the time.
- Transcription and searching both run on this device.
The model is never told what time it is
It writes a numbered citation like every other tool here, and the numbers are rewritten into timecodes afterwards from the transcript itself. That order is what makes a timecode in an answer trustworthy: it cannot be invented, because the model never had one to invent from.
Passages are about fifteen seconds of talking each, with no overlap. The chunk size used for documents is several minutes of speech, and a citation onto one of those can only say "listen from the start". Overlap is dropped because two citations onto the same sentence would name two different moments.
A recording with nothing in it is the hard case
No speech means no passages, so a question naming that recording matches the other files and gets answered out of them, describing four passages from three other meetings as though they were its contents. Facts about the library that no passage can carry are given to the model separately, and the screen states them as well, because the model read the note and refused anyway.
An identical file added twice is refused rather than taken twice. Two copies are two documents to everything downstream, and since agreement across documents counts as corroboration, a second copy scored as a second source and the answer cited one recording as though it were two.
Where nothing relevant is found, the question never reaches the model at all and you are told so. A fluent answer assembled out of the nearest available passage is indistinguishable from a real one, and on a recording nobody wants to listen to again that is the failure with no way back from it.
How many recordings can I add?
Up to ten files and about forty minutes in total, with each file capped at half an hour.
Can I play the moment it cites?
Yes. The timecode is a position in the recording, played back in the page.
Support and billing: support@gosmartpdf.com. Operated by Mrityunjay Kumar, India. Payments by Dodo Payments, our merchant of record.
Loading the editor…