Improving audio embeddings with squeeze-and-excitation: Introducing SaEENet

In this paper, we propose SaEENet, a novel neural network architecture to generate richer embeddings based on those generated by a pre-trained WavLM-large model and a set of convolutional layers fed by MFCCs. We employ 1D depthwise separable convolutions and GRU layers, and, to the best of our knowledge, we introduce for the first time the use of squeeze-and-excitation (SE) blocks for audio embedding weighting. The use of SE allows the model to assign a higher or lower relevance to each embedding generated from small audio segments and thus discard information generated from voiceless segments or segments with non-relevant information. In addition, we evaluated three different approaches for SE blocks to determine the most useful for the selected tasks. SaEENet outperforms similar models, such as the MEWHEV model, in the language identification, accent identification, and speaker identification tasks, achieving an improvement of 0.9%, 1.41%, and 4.01%, respectively, using 31.73% fewer trainable parameters. The results presented show that individual embeddings have varying effects on performance and that the integration of weighting mechanisms in SaEENet enhances accuracy in several speech classification tasks, highlighting the value of this approach for future applications.