Speaker verification using neural responses from the model of the auditory system / Noor Fadzilah Razali

Noor Fadzilah , Razali (2014) Speaker verification using neural responses from the model of the auditory system / Noor Fadzilah Razali. Masters thesis, University of Malaya.

PDF (Thesis (M.A)
Download (1789Kb) | Preview


    Speaker verification is the process of authenticating a person’s identity. Most of the available speaker verification systems have been implemented today is based on the step by step analysis of the acoustical signal itself. However, they are very sensitive to noise and work only at very high signal to noise ratio (SNR). On the other hand, the neural responses under noise are very robust, and the behavioral responses are also robust under diverse background noise. Therefore, a speaker verification system is proposed using the neural responses at the level of the auditory nerve (AN). For this, a very well-developed AN model by Zilany and colleagues (Zilany et al. 2009) is employed to simulate the neural responses on verifying a speaker. For this project, the feature extraction of the speech is analysed using the responses from the AN model, where the output is in the form of synapse output. A neurogram is constructed from the synapse responses of neurons with a wide range of characteristic frequencies. The neurogram’s average discharge or envelope (ENV) is then calculated. The resulted vector is then used to train the system using Gaussian Mixture Model (GMM) classification technique. Features are then extracted for testing data set and compared to the vectors for each of the trained speakers in order to verify a particular speaker. The speaker database is made up of recordings in a quiet room of 10 speech samples with 8 kHz sampling rate from 39 different speakers. Out of them, 70% speech samples of the speaker are used as the training set and the remaining 30% are for testing. As the neural responses are very robust to noise, speaker verification using AN model responses can substitute or outperform the current technology and thus improve performance for application such as in security processing.

    Item Type: Thesis (Masters)
    Additional Information: Thesis (M.Eng.) - Faculty of Engineering, University of Malaya, 2014.
    Uncontrolled Keywords: Speech perception; Computer simulation; Automatic speech recognition; Neural networks
    Subjects: T Technology > T Technology (General)
    Divisions: Faculty of Engineering
    Depositing User: Mr Prabhakaran Balachandran
    Date Deposited: 25 Oct 2017 16:25
    Last Modified: 25 Oct 2017 16:25
    URI: http://studentsrepo.um.edu.my/id/eprint/7784

    Actions (For repository staff only : Login required)

    View Item