Speech recognition in noise by using word graph combinations

In the practice, the performance of speech recognition systems is affected by speech signals being corrupted with various background noises in the environment. In this paper, we propose a new word graph combination (WGC) approach for speech-in-noise recognition. The aim of this work is to develop a method that would ensure robust speech recognition under various noise conditions, and in particular, under the adverse effect of environmental and impulsive noise. For this purpose, we developed a word graph combination (WGC) technique in which both continuous-mixture hidden Markov models (CMHMMs) and discrete-mixture hidden Markov models (DMHMMs) are being used as acoustic models. It has been previously verified that a DMHMM-based system can ensure significant improvements in the speech recognition performance under impulsive noise conditions. We also showed that the CMHMM-based system indicated better performance in high SNR conditions and environmental noise conditions. On the grounds of the above mentioned findings, we adopted a system combination approach in which both a DMHMM and a CMHMM are used. With the proposed method, complementary effects can be anticipated because the CMHMM and the DMHMM exhibit different error trends. Among the existing combination methods, which include recognizer output voting for error reduction (ROVER) and confusion network combination (CNC), in our work, we selected the technique of WGC. Unlike conventional combination approaches, like ROVER and CNC, the timing information for all word hypotheses is well preserved in the WGC. In the speech recognition experiments we performed, the proposed system showed better performance than the ROVER-based system or the baseline system. In particular, this new system showed comparatively higher performance under mixed noise conditions.