Repository navigation
v0.5.1
DemoLens 0.5.1 fixes the terminal windows that kept opening and closing on Windows, makes touch screens and pens work, reworks auto zoom so it is calm and never crops too tight, makes sure automatic cutting never removes something happening on screen, copes with quiet recordings, and fixes the silent editor preview.
The Windows and touch code here compiles and is exercised on CI, but it has not been run on real touch hardware. See "Check it on your touch laptop" and "Honest limits".
What was wrong on Windows
- Console windows. DemoLens is a windowed program, and every ffmpeg, ffprobe and helper it started opened its own console window for a moment, then closed it. Recording, the editor preview, export and the window list all start many of them. Every child process now starts through one helper that asks Windows for no window. A test fails the build if any code starts a process any other way, so it cannot come back on any OS.
- A PowerShell per refresh. Listing windows ran PowerShell each time. It now asks Windows directly, which takes microseconds and opens nothing.
- Screen scaling. On a 125 percent or 150 percent display the screen was captured smaller than the pointer positions that were logged, so clicks and zooms landed off target. The capture now uses real pixels, and the capture region is the real screen or window rectangle, including a second monitor to the left or above.
- Microphone. The Windows capture asked for a device literally named "Microphone", which most PCs do not have, so recording with a mic failed. It now uses the first real input device, or the one you choose.
- Touch. A tap on a touch screen does not move the cursor or press the left button as far as the old input watcher could tell, so taps were never seen.
Touch and pen
- Taps are clicks. Windows (Raw Input) and Linux on X11 (XInput 2) report finger and pen contacts. A tap or a press is logged as a click, so auto zoom and click effects treat it like one, and the key overlay shows "Tap". A drag with a finger is a drag.
- A tap ripple of its own. A soft fingertip disc and one ring, instead of the mouse ripple. It uses your ripple color.
- No double clicks. Apps that turn a tap into a mouse click would have counted it twice. Those duplicates are removed when recording stops.
- Editor on a finger. The timeline: drag, trim an edge (a larger grab area for a finger), double tap on the zoom lane to add a zoom, pinch with two fingers to zoom the view and slide to pan. Buttons, tabs, fields, sliders and the area picker handles are bigger on touch screens, and the 300 ms tap delay and double tap page zoom are off. An interrupted touch (a palm, a system gesture) cancels a drag cleanly.
Auto zoom, reworked
- Gentle by default. The default zoom is 1.5x and never goes deeper than 2x. A zoom keeps a margin of about 17 percent around what you clicked, so the target is never at the edge of the picture. A click right at the screen edge gets a gentler zoom still.
- Only when it means something. A zoom needs a cluster: two or more separate clicks close together, typing, or a marker you placed. A single stray click, or a double click, no longer zooms. Clicks far apart on the screen are split into separate zooms instead of one wide, useless zoom.
- Smooth. Zooms ease in and out in about a second. The camera follows the cursor only when it leaves a dead zone, and never leaves the screen. When you cut something out in the middle of a zoom, the camera now shortens the zoom instead of snapping, and zooms that touch each other animate every frame.
- Zoom intensity in the Auto edit panel: Subtle, Normal or Strong.
- Edit every zoom. Drag a zoom on the timeline to move it, drag its edges to resize it, change its level, click "Set focus point" and then click the preview to say where it should look, delete it, or add one at the playhead.
Cutting never removes screen activity
- Silence cutting is off by default. When you turn it on, a pause is cut only where the microphone is quiet and no key or mouse event happened and the picture did not change. Mic level alone is never enough, so a loud voice in the background cannot hide a pause, and a quiet moment while you are doing something is kept.
- You see the proposed cuts as red bands on the timeline first. "Apply cuts" makes them and "Discard" throws them away. Undo restores everything.
- The old short trims came from cutting on mic silence alone with only a short safety margin around input events and no look at the picture. We checked the splice and export code as well: picture and sound stay the same length across cuts (there is a test for that).
Quiet recordings and little or no audio
- A recording with a very quiet microphone is gated first and then lifted by at most 15 dB, so hiss is not turned into noise. A track that is only a noise floor is left alone.
- A recording with no sound track at all exports and transcribes without failing.
- Transcription says why it made no captions ("no sound", "too quiet", "no speech heard") instead of failing, and adds no captions when there is no speech. Stock phrases that speech models invent on silence are dropped. Existing captions are never cleared by an empty result.
The editor preview has sound again
- The sound has its own small file, built in seconds with the loudness evened out, and is played from memory. Before, the editor was silent until the whole 720p video copy finished building. If the first way fails there is a second way, and if both fail the editor says so in plain words.
- New volume and mute next to Play (key M). They change only what you hear while editing, never the export. If Mute is on under Style > Audio, the editor tells you the export will be silent.
- New AUDIO lane with the waveform, laid out through your cuts and speed changes.
- A browser test plays the real editor with a real recording and checks the clock moves, the player is unmuted, and sound actually comes out. It also covers: the fallback path, a broken sound file (it says so), and a recording with no sound.
Captions checked on real speech
- A test speaks four sentences with espeak-ng, transcribes them with a real model on CI, and checks the words, the caption times, that SRT and VTT agree, that the captions are saved with their word timings, that cutting filler words keeps the later captions in step, and that a recording with no speech gets no captions.
- Another test checks that the captions burned into the picture and the SRT show the same text at every moment, through cuts and a 2x speed change.
A clearer editor
- Every lane is labelled (CLIPS, ZOOM, NOTES, BLUR, KEYS, SPEECH, AUDIO). Empty lanes say what to do.
- Selecting a zoom, note or blur on the timeline opens its settings.
- The shortcuts list (press ?) includes the new keys.
- Screenshots of the editor are produced by CI on every change.
Other
demolens doctornow works on Windows and reports touch support, the virtual screen size, the windows it can see and the microphones it can see.- Quitting the app stops any recording, meter or decoder ffmpeg that is still running.
- CI records the real Windows screen and a region, with spaces in the folder name, then builds a project and exports it, and checks that no ffmpeg is left running.
- The release build now checks that the bundled Windows ffmpeg has gdigrab and dshow.
Check it on your touch laptop
- Update from 0.5.0: Settings > About and updates > Check for updates.
- Start a recording and watch the taskbar. No console window should flash.
- Run
demolens doctorin a terminal. Undertouchit should say your screen is supported. Send me the output if it does not. - Record 10 s tapping the screen with a finger, then open it. You should see a tap ripple and an automatic zoom at each tap.
- In the editor try a pinch on the timeline and a double tap on the zoom lane.
Honest limits
- Voice isolation is not included. Removing a loud voice in the background from the microphone needs a separate model. "Remove background noise" still removes steady noise only. A loud background voice can still be transcribed. We only make sure it never decides where to cut.
- The preview sound was tested in Chrome with a real recording. The macOS and Windows web views are not available to the author, so if sound is still missing there, the editor now shows why in the sound label next to the volume.
- Touch capture is checked only by compiling it for Windows and Linux. Raw Input on some touch drivers may not report contacts. If it does not,
doctorsays so and I need your device name. - Touch positions map to the main monitor. A rotated screen is not handled.
- Linux touch needs X11 with XInput 2.2. Wayland is not covered.
- Windows system audio capture is still not available. The microphone is.
- macOS has no touch screens, so nothing changes there except the process clean-up and the editor fixes.
- macOS capture is still not run on real hardware by the author.
- Caption word times from the small speech model are rough. A held word is shortened and words piled up at the end are spread over the time the voice is really on. A bigger model gives better timing.
- The speech engine is now built without CPU specific instructions, so the installers do not crash with "illegal instruction" on a processor older than the one that built them.