Adaptive speech model for missing-feature reconstruction

摘要

This paper presents a new adaptive speech model for Missing-Feature Reconstruction using unsupervised learning for speech recognition. Hence, a neural network with time-varying structure, LARFSOM, and a FNNS algorithm to find two best matching units were used. For evaluation purposes, Aurora 2 and NOIZEUS databases were used. Experimental results indicate that the model is robust to noise without Oracle knowledge.