Design And Implementation Of Embedded Syllable Recognition System

In the rapid development of human-computer interaction technology today,speech recognition play a more and more important role.The need of speech recognition is not only for entertanment.Speech recognition plays a more important role in solving the real-time,convenience,intelligence and other issues of human-computer interaction.Among the speech recognition technology,the recognition of isolated vocabulary after decades of development,has gained fruitful results.So far,the recognition of isolated vocabulary has gained wide applications in voice command control of Mobile robot,voice command control of car,smart home system.The graduation thesis has a research on a embedded vocabulary voice command control system based on DSP.The thesis studies the hardware and software scheme of speech acquisition from the bottom.In the thesis a hardware and software system is set up with TMS320C5509 A as the core hardware.While collecting voice,the system detect the endpoint of voice signal.The thesis gives a brief introduction to the signal preprocessing process and the signal feature extraction scheme which are commonly used in speech recognition.Combined with the recognition of the number of entries and the processor computing ability,the thesis makes a brief exploration of speech recognition primitives.And inspired by the most widely used three-tone model of speech recognition technology,the thesis uses syllables as recognition primitives.In order to narrow the number of entries while matching with templates,the thesis classifies the syllables in the speech library.Clustering is based on the ratio of the energy of the initials and vowels,the pitch period of vowel and other parameters.The thesis gives an introduction to the algorithm of the segmentation of initials and vowels and pitch period estimation algorithm in clustering process.The thesis implements the algorithm.The thesis adopts Mixed Gaussian Hidden Markov Model to match the syllables.In the thesis,the meaning of model parameters and methods for calculating parameters of the models are introduced.Based on this,a fast gaussian algorithm aiming at reducing the amount of calculation is introduced.The embedded system in the thesis usese the built-in digital signal processing library functions to optimize the amount of calculation.The functions contain correlation calculation function and matrix calculation function.When the system works on 100 control commands and 100 phrases,the final recognition reaches up to 92% within 5s.