Speaker Verification for a Telehealth System

Project Type: Master/Diploma Thesis
Student: Schönemann Nikolaus Gerhard
Mentor: Gernot Kubin

 

 The goal of this work is to develop a speaker verification system and finally to build a demonstrator, which is intended to be integrated in a Zydacron telehealth system. Speaking of speaker verification, we mean to verify a speaker's identity by his voice. In such a system, the verification of the speaker's identity happens online during a doctor-patient conversation. Our verification system is text-independent and based on statistical Gaussian mixture models (GMM). We simulate the system in MATLAB using several approaches, such as different types of features (LPCCs and MFCCs), number of Gaussian mixture components and speaker model adaptation methods. The system's performance is evaluated with clean and noisy speech utterances using the WSJ0 database. For clean speech, our simulated system reaches a performance of 0.09% equal error rate (EER). We also found out, that a close/far speaker decision can be made by using special trained models for comparison. Next, a system demonstrator is implemented on a TI6416 fixed-point DSP in combination with a PC host program, using the parameters that yield the best simulation results. For the fixed-point implementation we reach an EER of 2.95% for clean speech. Eventually, the fixed and floating-point implementation are compared, and fixed-point specific problems and the associated loss in the system's reliability are discussed.