Video Timestamps
Make your media library readable.
Video and audio files are usually black boxes. Dhito transcribes them in the background and captions what appears on screen, letting you search a recording by what was said or shown and jump to the exact timestamp.
Why your recordings are invisible to search
macOS indexes a video file by its name and its standard metadata — duration, codec, creation date. The spoken content is not part of that index, so a recording named zoom_0.mp4 is, as far as search is concerned, a file with no content at all.
For anyone who records meetings, interviews, lectures, or research calls, that means the most information-dense files on the disk are the least findable. You end up scrubbing through a timeline from memory, looking for the ten seconds that mattered.
Transcription turns media into searchable text
Dhito runs OpenAI's Whisper model locally on your Mac to transcribe audio and video in the background as it indexes. The transcript is then embedded into the same semantic index as your documents, with timing information preserved for each segment.
Because the transcript is indexed semantically rather than as plain keywords, you do not need to recall the exact phrasing. Searching "pricing discussion with the marketing team" can surface a call where nobody said the word "pricing" — they talked about rate cards and margins, and the model places that near the same meaning.
The timing data is what makes the result useful. Dhito returns the moment rather than the file, so a hit in a two-hour recording opens at the 47-minute mark instead of at the beginning.
The picture is indexed too, not just the audio
Transcription only covers what was said, which leaves out a lot: the slide that was on screen, the chart someone shared, the figure nobody read aloud. So Dhito also looks at the video itself. As it indexes, it samples frames across each segment of the recording, and for every frame it writes a caption of what is shown, notes the objects it recognises, and OCRs any text visible on screen.
Each of those is stored against the timestamp it came from, in the same semantic index as the transcript. A search for "the slide with the Q3 revenue breakdown" can land on the minute that slide was displayed, even in a recording where nobody said the words out loud.
Where on-screen text is dense or stylised and the vision model reads little of it, Dhito falls back to macOS's own OCR engine for that frame. Both run on your Mac, and the video is never uploaded.
Local Transcription
Using the powerful OpenAI Whisper model running natively on your Mac, Dhito converts spoken dialogue from Zoom recordings, podcasts, and video tutorials into searchable text.
Search by Spoken Word
Remember a specific quote from a 2-hour meeting but don't remember when it was said? Search for the exact phrase and Dhito will find the video.
Jump to Timestamp
Dhito doesn't just find the file; it provides you with the exact timestamp where the phrase was spoken, so you can start watching right at the relevant moment.
Common questions
Which video and audio formats are supported?
17 video containers and 14 audio containers. That covers the common ones — MP4, MOV, MP3, M4A and WAV — as well as MKV, WebM, AVI, M4V, MPEG and MXF on the video side, and FLAC, AAC, OGG, OPUS, AIFF and ALAC on the audio side. Screen and meeting recordings exported from the usual conferencing tools are all included.
Can I find a moment by what was on screen rather than what was said?
Yes. Alongside the transcript, Dhito samples frames through each segment of a recording and captions them, records the objects it recognises, and OCRs on-screen text — all tied to the timestamp they came from. So a slide, a chart or a screen-shared document can be found even if nobody described it aloud. Silent screen recordings and screencasts are searchable for the same reason.
Does transcription happen in the cloud?
No. Whisper runs on your Apple Silicon, so recordings are never uploaded. This is the difference that matters for confidential meetings: the audio does not leave the machine at any point.
How long does transcription take?
It runs in the background and depends on the length of the recording and your chip. You can keep working while it processes, and files become searchable as each one finishes rather than only when the whole queue is done.
Can I search for a topic rather than an exact quote?
Yes — that is the point of indexing transcripts semantically. Exact phrases work, but so does describing the subject, because matching happens on meaning rather than on the literal words spoken.
What happens with speakers who overlap or have strong accents?
Whisper is robust across accents and handles conversational audio well, but no transcription is perfect. Semantic search softens the impact: a query still matches the surrounding context even when an individual word came through wrong.