Connected Digit Speech Recognition Based On The Acoustic Universal Structure

Connected digit speech recognition has wide application in our daily life, such as speech autodialer in telephone and telecommunications systems, personal number verification, intelligence home appliances'remote control for air conditioning and television channel, etc.Connected digit speech recognition is a crucial branch of continuous speech recognition. The following are the difficulties that connected digits speech recognition is facing: First, it's uncertain about the length of connected digits and difficult to determine the boundary of each digit unit accurately; Second, connected digit strings are the combination of digits without any limitation and grammatical knowledge; Third, the digit speech pronunciation has its own characteristics, for example, high degree of confusion between some digits, the phenomenon of co-articulation and so on. In addition, speech acoustics is inevitably distorted by non-linguistic features. All these factors make the performance of the recognition system poor.This paper mainly focuses on two aspects:(1) Connected Digit Speech Recognition Based on the Acoustic Universal Structure. This paper applied a novel acoustic representation of speech to connected digit speech recognition, called the Acoustic Universal Structure where the non-linguistic variations such as vocal tract length, lines, noises, etc, are well removed, and proposed a double-layer structure of speech model matching strategy. The recognition system can realize variable length connected digit recognition in the absence of grammatical knowledge, a large number of training template and universal channel normalization. So the robustness problem for speaker variations which exists in current speech and connected digit recognition system is well solved.(2) Robust Speech Recognition Based on Histogram Equalization. The additive noise is also an important reason that weakens the performance of recognition system. Comparing with classical methods, the histogram equalization can reduce non-linear distortion and improve the robustness of speech recognition system quite well. However in many applications, the feature distribution between training and test speech is usually not identical because of their difference in phonetics or acoustics, then the validity of HEQ can be weaken. This paper proposed Histogram Equalization of Classified Features algorithm, the experiments show that this method improves the performance of system with comparison of usual histogram equalization.