TL;DR
This paper introduces SpeechMarker, a novel multi-level speaker verification (SV) system designed for attendance tracking. It combines voice-password, text-dependent, and text-independent SV modalities within a user-friendly GUI, offering a more robust and convenient authentication solution than single-modality systems.
What Problem Does This Paper Address?
Traditional attendance systems often rely on manual sign-ins, which are prone to errors and manipulation, or single-factor biometric systems that can be vulnerable to spoofing. While speaker verification (SV) offers a promising biometric alternative, existing systems often focus on a single modality, which can have limitations. Text-dependent SV is more user-friendly but susceptible to spoofing, while text-independent SV is more robust but can be less intuitive. This paper addresses the need for a more secure, user-friendly, and practical attendance system by developing a multi-level SV framework that leverages the strengths of different SV modalities.
Methods
SpeechMarker employs a three-tiered approach to speaker verification:
- Voice-Password: Users enroll with a unique, fixed phrase. Verification uses Mel Frequency Cepstral Coefficients (MFCCs) and Dynamic Time Warping (DTW) for temporal alignment.
- Text-Dependent: All users share a common phrase for enrollment and testing. Similar to the voice-password module, it uses MFCCs and DTW. To mitigate spoofing, a set of three fixed phrases are used during enrollment, and one is presented randomly during testing.
- Text-Independent: This module utilizes an i-vector based speaker modeling approach. Training involves about a minute of speech, and testing uses short phrases. LDA and WCCN are used for channel/session compensation.
The system's decision can be made by any individual module or by fusing their results. A graphical user interface (GUI) is provided for both user enrollment/attendance marking and a "debug mode" for administrators to analyze system behavior.
Datasets
The paper mentions the use of the NIST SRE 2010 database for learning background models for the text-independent module. No new datasets were introduced.
Key Insights and Ideas
- Multi-level Fusion for Robustness: Combining different SV modalities (voice-password, text-dependent, text-independent) offers a more robust and secure attendance system than relying on a single method.
- User-Centric Design: The development of a user-friendly GUI with both "user" and "debug" modes caters to both end-users and system administrators, facilitating practical deployment and analysis.
- Addressing Spoofing: The text-dependent module incorporates a mechanism (randomly presenting one of three phrases) to add a layer of defense against spoofing attacks.
- Practical Deployment: The system has been successfully deployed in a laboratory environment for a closed group, demonstrating its real-world applicability.
Key Results
The paper highlights the successful deployment of the SpeechMarker framework for regular attendance marking within a laboratory setting. While specific quantitative performance metrics (e.g., accuracy rates for each modality or the fused system) are not detailed, the successful deployment itself serves as a key qualitative result. The paper emphasizes the system's ability to provide a practical and user-friendly attendance solution.
Limitations and Future Work
The paper acknowledges that the system has been deployed among a "closed group in a laboratory environment." This suggests that its performance and robustness in larger, more diverse, and uncontrolled real-world scenarios may require further investigation. Future work could involve scaling the system to larger user bases, evaluating its performance under various acoustic conditions and with different types of microphones, and exploring more advanced fusion techniques for improved accuracy and security.
Summary updated