Orthographic Comparison Revealed by Ambient Sentiment Classification

Several deep neural network versions have been successfully used to create parametric models that project variable-duration spoken word segments onto fixed-size vector representations, also known as acoustic word embeddings (AWEs). However; it is uncertain to what extent the distance in the growing AWE space can evaluate word-form similarity. This study investigates whether acoustic embedding space distance is related to phonological dissimilarity. The performance of supervised techniques for AWEs with different neural architectures and learning aims to solve this question. AWE models are trained in controlled settings for two languages (German and Czech) and evaluate the embeddings on two tasks: word discrimination and phonological similarity. The findings reveal that (1) the embedding space distance weakly corresponds with phonological distance in the best circumstances, and (2) enhancing word discrimination performance does not always result in models that better reflect word phonological similarity. The results emphasize the need to reconsider existing intrinsic AWE evaluations.