Recognition of Speech with Non-random Attributes

Speech recognition, Hidden Markov Models, HMM

Recognition of Speech, where utterances are supposed to have assigned attributes with Non-random character

Most of current speech recognition systems are based on Hidden Markov Models assuming that speech features are sequence of stationary stochastic processes. However, there are certain speech attributes, such as background noise type or speaker voice color, that do not have stochastic character. This fact is often ignored, by designers of robust speaker independent recognition system. In this work, we investigate how the performance of a noisy speech recognition can be improved provided that we have prior knowledge about type and level of noise. Next, recognizer that is using separate models, each trained on a particular type and level of noise, is proposed for more appropriate modeling of speech.