Modeling a speech signal

For modeling the time series of the spoken word manna we used a similar network compared to the saxophone model. Due to the increased instationarity in the signal we needed an increased number of RBF units in the network. The best results up to now has been obtained with a network of 400 hidden units, delay time , output dimension 8 and input dimension 11.

In figure  we show the original and the resynthesized signal. The quality of the model is not as high as in the case of the saxophone. Nevertheless, the word is quite understandable. From the figure we see, that the main problems stem from the transitions between consecutive phonemes. These transitions are rather quick in time and, therefore, there exists only a small amount of data describing the dynamics of the transitions. We assume that more training examples of the same word will cure the problem. However, it will probably require a well trained speaker to reproduce the dynamics in speaking the same word twice.

  
Figure: Original and synthesized signal of the word manna.