The Research And Implementation Of Embedded Speaker Recognition System

Speaker recognition is to find out the speaker's identity based on the spoken utterances. As one of the most important biology authentication technologies, it can be widely used in many security areas. In this work, we'll build a speaker recognition system based on the platform of ARM9, using Gaussian mixture models-universal background model. We focus on three key parts: feature extraction, speaker recognition's algorithm, and hardware platforms.Feature extraction is generally carried out to transform speech waveform into a sequence of observation vectors. Currently, the predominant choice of parameters is mel frequency cepstral coefficient (MFCC). In this thesis, we will describe the extraction of MFCC and its optimization.For the speaker recognition algorithm, we firstly introduce some widely used algorithms: hidden Markov model (HMM), Gaussian mixture models (GMM), and artificial neural networks (ANN). In recent years, it is more common to represent speakers with the Gaussian mixture model (GMM). We'll address in detail three issues: uniform background model (UBM), the adaptation of speaker model, and the verification strategy.For the hardware platform, we firstly explain the reason to select ARM9 H2410EB. Then we describe the design and implementation of speaker verification system based on this embedded platform. Several new calculational methods are applied to improve the verification performance, such as the transformation from floating-point operation to pointing operation, pretreatment of the calculation, fast and approximate mathematic calculation, and MAX mechanism etc. These methods guarantee the real-time performance of embedded system.