Publisher's Synopsis
In theory multiple languages can be recognized just as one language. However, current state of the art speech recognition systems are based on statistical models with many parameters. Extending such models to multiple languages requires more resources. Therefore a lot of research in the area of multilingual speech recognition has proposed techniques to reduce this need for more resources through parameter tying across languages. This work shows that tying at the density level of Hidden Markov Model based speech recognizers offers the greatest flexibility for the design of a multilingual acoustic model. Furthermore, new algorithms are designed and tested for a fast and efficient creation of systems for many different language combinations. These algorithms base on the addition of only relevant Gaussians and on the projection of a Gaussian mixture distribution to new sets of Gaussians. The positive aspects of the architecture proposed in this work are that non-native accent recognition fruitfully applies knowledge about the mother language of the speakers and that an optimal resource allocation for each language can be guaranteed through an online adaptation to the current tasks.
