FrameQuery 1.1.1: Scene Search and Face Sampling Fixed

FrameQuery 1.1.1 is out. It fixes a bug where search returned only filenames and video pages showed no scenes, makes the per-second face detection introduced in 1.1.0 actually run, and adds voice matching for transcript lines that were never assigned a speaker. Free update for all users.

Find the moment again, not just the filename

If your searches had started returning only file names, and opening a video showed an empty scene list, that was a locking bug in the search index. When the index could not be opened for writing — most often because a second copy of the app was still shutting down — the whole index went unavailable rather than falling back to read-only. Nothing told you. Search quietly dropped to filename matches until you restarted.

The index now serves reads even while another process holds the write lock, and reclaims that lock on its own when it frees. No restart, no lost results.

Search also returns several moments per video now instead of collapsing each video to a single hit. Search a person who appears throughout an interview and you get their separate appearances with their own timecodes, rather than one row pointing at the first one.

Per-second face detection now actually runs

Version 1.1.0 added the option to check a frame every second instead of one frame per scene. The frames were being extracted and then thrown away: the detector needs a video's frame rate to place them on the timeline, and when that had not been recorded it discarded the lot and silently fell back to one frame per scene. On one library that was 24,684 extracted frames going unused.

The frame rate is now recovered by probing the file when it is missing. Detection sees the full sampled set, so people who only appear between scene changes get found.

Two related fixes come with it. Faces detected in sampled frames were being grouped as a separate person per frame, so the same face appeared over and over in the People tab — they now group correctly. And their boxes are no longer drawn over scene thumbnails they were never measured against; they appear as clickable timecodes under the scene instead, which is where they actually happened.

Face and voice processing continues to run entirely on your machine. See how people matching works and what stays local.

Put a name to speakers the transcript missed

Roughly half the transcript lines in a typical library carry no speaker, mostly from footage processed before speaker diarization existed. Where a video has some speakers already identified, FrameQuery now matches the unlabelled lines against them by voice — same recording, same microphone, so the comparison is reliable.

A line is only claimed when one speaker is a clear match and clearly better than the runner-up. Anything ambiguous stays unlabelled rather than guessing, and attributed lines are marked in the transcript so you can tell them from the ones diarization produced. Re-process voices to apply it to your existing library; new footage does it automatically.

Assigning people got quicker too. The assign dialog now has a search box and can create a person inline, so you no longer have to cancel out, make the person, then find the face again. You can also right-click a speaker tag in the transcript to assign it.

What's fixed

Search returning only filenames, and video pages showing no scenes, after the index failed to open
The app crashing on transcripts containing accented characters
A single failed frame upload discarding an entire completed upload — frames now retry
Interrupted uploads holding processing slots until you hit the "too many active jobs" limit
Upload progress freezing short of 100% instead of completing
Face boxes drawn in the wrong place on scene thumbnails
Transcript lines with no speaker being labelled "Speaker 1" as though they were one

The app updates itself, and your library, projects, tags and people are untouched. Search results come back on their own; picking up the extra faces and the newly named speakers needs Re-process faces and Re-process voices respectively, both optional and both local. FrameQuery 1.1.1 runs on Windows and Apple Silicon — see supported formats and systems, or download the latest build.