Cognitive science shows that perceptual task, such as similarity measurement, is carried out in a low dimensional representation space, and patterns are represented in a topological view in the low dimensional space by considering their similarity relationship. Accordingly, it is assumed that speech production and perception in the high level are carried out in lower dimensions with a similar topology. In this paper, we used an unsupervised learning method, i.e., Locally Linear Embedding (LLE) to extract low dimensional structure of five Japanese vowels using articulatory data with eight observation points. The results showed that the learned topological structure in articulatory space with a low dimension is consistent with F1-F2 pattern of the vowels in acoustic domain. Because the acoustic data is produced by the articulatory movement which is controlled in motor area in brain, these kinds of topological structures may suggest the cognition of speech in high level by invariant topology mapping between different spaces. Also, from neural mechanism for pattern encoding aspect in high level, with the evolution of neurons which are exposed to so many patterns, the plasticity of neurons is adapted to encode all speech patterns efficiently by reference encoding which explores their similarity relationship between them.
展开▼