Implementation and analysis of GMM-based speaker identification on FPGA

Abstract

The use of highly accurate identification systems is required in today’s society. Existing systems such as pin numbers and passwords can be forgotten or forged easily and they are no longer considered to offer a high level of security. The use of biological features (biometrics) is becoming widely accepted as the next level for security systems. One of the biometric is the human voice and it leads to the task of speaker identification. Speaker identification is the process of determining whether a speaker exists in a group of known speakers and identifying the speaker within the group. Speaker specific characteristics exist in speech signals due to different speakers having different resonances of the vocal tract. These differences can be exploited by extracting Mel-frequency Cepstral Coefficients (MFCCs) from the speech signal. A statistical modelling process known as Gaussian Mixture Model (GMM) is used to model the distribution of each speaker’s MFCCs in a multi-dimensional acoustic space. GMM involves with two phases called training and classification. The training phase is complex and is better suited for implementation in software. The classification phase is well suited for implementation in hardware and this allows for real time processing of multiple voice streams on large population sizes. Several innovative techniques are demonstrated which enable hardware system to obtain two orders of magnitude speed up over software while maintaining comparable levels of accuracy. A speedup factor of eighty six is achieved on hardware-based FPGA compared to a software implementation on a standard PC for this approach.