Gentle Reminder: Reg CSE Speech Seminar

[Mari Ganesh Kumar M]

Dear All,
A gentle reminder on  speech seminar 

​

scheduled today.

Regards,
Mari Ganesh Kumar M.

On Mon, Mar 6, 2017 at 9:38 AM, Mari Ganesh Kumar M <marigan...@gmail.com> wrote:

Dear all,

The following speech seminar has been

​​

scheduled tomorrow.  Please make it convenient to attend.

Title: A new paradigm for speaker recognition and its effect on spoofing attacks

Name: Saranya M S (CS15D006)

Location: BSB 361 (Seminar Hall)

Time/Date: March 7, 2017, 2.00 - 3.00 pm

Abstract:

Speech technologies today have become ubiquitous owing to the small form factor of smartphones. Interactive Voice Response (IVR) systems are beginning to get replaced by speech recognition based interactive systems. Speaker recognition aims to analyze and extract information from the speech signal to recognize the speaker’s identity. Conventional speaker verification systems utilize information from different feature representations using fusion. The first part of the research is to propose an alternative technique which achieves a similar effect but utilizes a more effective feature selection technique with less computational cost. The underlying assumption of the method is that different speakers are better represented, and hence better verified, in different feature spaces. This technique, which we term as feature-switching, performs verification using a feature representation most suitable to the speaker under consideration. Out of a possible set of candidate representations, the most optimal representation for a speaker is determined during training. Verification is performed using the optimal feature of the claimed speaker. Determining the candidate feature for every speaker is an important task. Effect of feature-switching will be shown utilizing the classical GMM-UBM speaker verification system, as well as state-of-the-art i-vector-based speaker verification system. The performance of speaker recognition systems is measured based on the robustness of the systems against an impostor speaker. Speaker verification systems are robust against impostor attacks, nevertheless, they are vulnerable to spoofing attacks. The spoofing attack can happen at any stage of a speaker recognition system right from sample acquisition stage to decision-making stage. Few common spoofing attacks are (i) impersonation (ii) replay attack (iii) speech synthesis and (iv) voice conversion. All these attacks happen at the microphone level or transmission level. Owing to the small form factor of the smartphones, inexpensive and high-quality recording and playback devices replay attacks pose a major threat to the ASV system. This replay attack can be performed without any speech processing knowledge. The second part of the research is to give a countermeasure for this replay attack.

Regards,
Mari Ganesh Kumar M.

ᐧ

ᐧ