How to interpret the output of the segmentation model ?

[amitli1]

pyannote for speaker diarization based on the following segmentation model:
End-to-end speaker segmentation for overlap-aware resegmentation

In the above paper they wrote, under the Implementation details:

• model input: sequences of 80000 samples
  [i.e: 5s audio chunks with a sampling rate of 16kHz]
• model output:
  K max -dimensional speaker activations between 0 and 1 every 16ms.

1. Does it means that the output shape is (K, 5000/16) ?
2. The output values are between 0 and 1. how to interpret it ?
   How to conclude if we have a new segment or number of segments in each output ? number of speaker in output ? (example will be very helpful)

[hbredin]

Did you read this? This should answer most of your questions about this model.

[amitli1]

Thanks,
Still don't understand the shape: (11, 293, 3) for 5s sliding window.
Didn't understand the value of 293.
5000 / 16ms = 312.5
How they got 293 ?

[amitli1]

@hbredin Can you please explain how we get the shape of: (11, 293, 3)
after the code:

inference = Inference(model, duration=5.0, step=2.5)
output = inference(SAMPLE_WAV)

I don't understand why 293 ?

[benbot]

@amitli1 did you ever figure it out?