Robust Parallel Speech Recognition

This book focuses on automatic speech recognition in clean and noisy or reverbrant environments. Therefore, a parallel speech recognition system using TempoRAL Patterns (TRAPs) is described. The TRAPs are computed over a rather long temporal context for each critical band in the signal's spectrum. Then, the features of the different bands are combined. Thus recognition only in certain bands is possible. This is beneficial if noise only occurs in parts of the spectrum. In this manner multiple speech recognizers are trained which analyze disjoint parts of the frequency domain. Each of the speech recognizers extracts a different word chain from the audio signal. In the end the word chains are merged to form a single recognition result. As shown on different data sets the parallel speech recognition system is much more robust to noise and reverberation than the state-of-the-art baseline system.