A recurrent neural fuzzy network for word boundary detection in variable noise-level environments

This paper discusses the problem of automatic word boundary detection in th e presence of variable-level background noise. Commonly used robust word bo undary detection algorithms always assume that the background noise level i s fixed. In fact, the background noise level mag vary during the procedure of recording. This is the major reason that most robust word boundary detec tion algorithms cannot work well in the condition of variable background no ise level, In order to solve this problem, we first propose a refined time- frequency (RTF) parameter for extracting both the time and frequency featur es of noisy speech signals. The RTF parameter extends the (time-frequency) TF parameter proposed by Junqua et al, from single band to multiband spectr um analysis, where the frequency bands help to make the distinction between speech signal and noise clear. The RTF parameter can extract useful freque ncy information, Based on this RTF parameter, we further propose a new word boundary detection algorithm by using a recurrent sell-organizing neural f uzzy inference network (RSONFIN). Since RSONFIN can process the temporal re lations, the proposed RTF-based RSONFIN algorithm can find the variation of the background noise level and detect correct word boundaries in the condi tion of variable background noise level. As compared to normal neural netwo rks, the RSONFIN can always find itself an economic network size with high- learning speed, Due to the self-learning ability of RSONFIN, this RTF-based RSONFIN algorithm avoids the need for empirically determining ambiguous de cision rules in normal word boundary detection algorithms. Experimental res ults show that this new algorithm achieves higher recognition rate than the TF-based algorithm which has been shown to outperform several commonly use d word boundary detection algorithms by about 12% in variable background no ise level condition. It also reduces the recognition error rate due to endp oint detection to about 23%, compared to an average of 47% obtained by the TF-based algorithm in the same condition.