Speaker-Embeddings-Correlation-Pooling
This is the original implementation of the pooling method introduced in "Speaker embeddings by modeling channel-wise correlations" by T. Stafylakis, J. Rohdin, and L. Burget (Interspeech 2021), a result of the collaboration between Omilia - Conversational Intelligence and Brno University of Technology (BUT), which you may find here.
The code is in TensorFlow1 (TF1) but it should work with TF2 too. I only provide the code for creating the network and the required hyperparameters. The training hyperparameters we used can be found in the paper.
The code is well-commented, at least the part and (hyper-)parameters required for the correlation pooling.
Apart from the experiments provided in the paper, the code allows the user to: (a) Combine standard statistics pooling with correlation pooling, by concatenating the two pooling layers into a single one, and (b) Extract correlation pooling from outputs of all 4 internal ResNet blocks (aka stages) and concatenate them in the pooling layer.
The code can be more efficiently written using tensor-only operators. However, to facilitate research we have implemented it using lists of tensors, e.g. after merging frequency bins to frequency ranges. Despite this inefficiency, we observe no differences between correlation pooling and standard stats pooling in training speed.
Start with the file train_resnet.py, which creates the ResNet (with the pooling mechanism) and sets its parameters. All parameters are set so that you reproduce our best performing experiment (P7 in the paper).
So, try it and let us know what you'll get! Themos