Subspace-GMM acoustic models for under-resourced languages: feasibility study

Acoustic model parameter estimation is hampered by a lack of data. To reduce the number of parameters to be estimated, we propose sub-GMM modelling, which constrains the acoustic models to a low- dimensional manifold embedded in the space of Gaussian mixture weights. The manifold model is obtained through non-negative matrix factorization with sparsity constraints. Our preliminary mono- lingual experiments show that the proposed model is as efficient as clustering the distributions to a smaller set, while it opens perspectives for a new parameter tying technique. In the example, the number of parameters to be estimated per distribution is reduced more than an order of magnitude.