jeremy110

You can use the code to save the picture. 

I revise the code.It works for me. 

I can reply your second question.When you run the speakerDiarization.py,you can change both of them in here. 

I don't have a great solution because each feature of audio isn't the same. I test 10 min news of Taiwan and then embedding_per_second=0.9 and overlap_rate=0.3 are absolutely great for...

In your first href.It's to produce weight.h5. 

You can directly reduce your batch size and increase the epochs.

If you know that the speaker number correspond to your speaker name,you can add your speaker name here in advance.

Let me make sure I have it right.So you want to revise the predict function and it can return your speaker name?

Maybe you can add your speaker name which cluster_id correspond to in BeamState in uisrnn.py and use it when prediction function return the results.

you must assure the shape of feats is right.For example:(256,512),the 512 is your observation_dim