Implementation Isolated-Word DSP Speech Recognition System Based On Dynamic Vocabulary

The speech recognition is an important topic in pattern recognition field, because its development will deeply influence the future of human-computer interface. Among the researches on the speech recognition, the embedded speech recognition based on dynamic vocabulary is still a challenging one. Although many embedded speech recognition systems have been built up in recent years, their performances is far from satisfying and good enough for the application on a large scale. Therefore, the further research on this topic is necessary and of great importance.Aim of this paper is to implement isolated-word speech recognition system using DSP. The isolated-word speech recognition module based on dynamic vocabulary is developed to simplify its training process, make it easier to add recognition content, and improve the isolated-word recognition ratio as well as the recognition speed. The application will be more convenience and human.Firstly, the fundmental theories of speech recognition and the applications of HMM in speech recognition are discussed.Secondly, the function requirments and structure design of isolated-word DSP speech recognition system based on dynamic vocabulary are intruduced. The development and characteistics of embeded DSP are summaried. Then the impementation scheme is constructed. The system models and the fixing-point scheme and programme optimizing scheme are discussed.The Training Model: The basic acoustic model is produced using HTK. The training set is a robust language model. The standard acoustic model document is produced. In this paper, the acoustic model which is trained by HTK is used.The Reference Model library Model: the lexicality, auto-label, and reference model library part are involved. The reference model is indexed by the label documents of the dictionary and is composed of the basic acoustic model dynamically. Because of the limitation of the DSP memory, the acoustic model is divided into two parts and loaded into DRAM orderly. The index is arranged in both the states' and the models' order. There are 61 models and 183 states and in large size. To solve this problem, the data is written to a fix memory space of the DSP.Speech Signal Processing and Recognition Model: the port detection, the character extraction, and other key processes are included in the speech signal processing. The recognition process measures the comparability of the speech signal eigenvector and the reference model.At last, The problems and difficulties durring the system implementation and their solutions are involved in this dissertation in which the future direction of the study is pointed.