Labels for silent detector

[henryxrl]

Labels for silent detector

[Toku11]

In your preprocessing you are creating a "bit mask" which is supposed to be your labels for the silent interval detection however your bitmmask is always 1 because you created it like '1' * file_info[FIELDS[5]], how is that you use it to train

[aloha0920]

Hi, same issue here. Did you figure it out yet?

[henryxrl]

@Toku11 You are correct. Each 1/0 bit indicates a 1/30 time interval. And yes, 1 means speech, 0 means silence. Each piece of training data is indeed 2 seconds which means 60 of those 1/0 bits.

In the preprocessor_audioonly.py file, the reason why it's setting all 1's is that this generates data for inference job, NOT for training (so all 1's doesn't matter - you'll get the correct 1/0 sequence as a result). For training, you'll have to provide the ground truth correctly labeled 1/0 sequence. You can get creative and create your own. My way of doing so is to look at the power of the speech signal and label each 1/30 time interval 0 (silence) if the power is below a certain threshold, and 1 if above. As a reference, you can take a look at the 'get_bitstream_better' function (commented out) from the util.py file in the preprocessing directory.