Say “Hey Siri” and the phone wakes up. That step is keyword spotting, or KWS. Kwscope is a test bench for wake-word models. Talk into the microphone or open a recording, and every time the model wakes, you see it on the spectrogram.
Kwscope 0.2.0 is open source under MIT OR Apache-2.0. Today it is what we use to test our wake-word models in house. Soon we will release an open-weight wake-word model you can load and try.
What you see is what the model decides
A wake-word model is judged one moment at a time. Did it wake when it should have? Did it stay quiet when it should have? Kwscope matches the training-side runtime frame by frame, so every trigger on screen is a decision the model would really make once it ships, not an approximation.
Project snapshot
Kwscope ships without a model. When the open-weight wake-word model is out, it will be linked from here.
Two stages
Stage one listens, stage two double-checks
Frame by frame
Matches the training-side runtime
0.2.0
Current release · MIT OR Apache-2.0
Every wake has a place
1/3
Green is a wake that was accepted. A red dashed box is a false alarm stage two caught. The number above each is stage two’s confidence. Below runs stage one’s score; the moment it crosses the yellow dashed line, the model wakes.
Kwscope 0.2.0 on a MacBook Pro, listening to the microphone: four triggers, three accepted, one caught by stage two.
Move a threshold, find where it belongs
2/3
Wake words always trade one failure for another. Too eager and it wakes on its own; too cautious and it ignores you. Drag t1 or t2 and the open recording is judged again at once, so you see which wakes appear and which ones vanish.
At the microphone, and in the pipeline
3/3
The best test is to speak. With the microphone on, the spectrogram scrolls as you talk and each wake is marked as it fires. For a batch of recordings, kwscope stream prints one JSON line per trigger, ready for your evaluation.