Beyond the Equal Error Rate: About the inter-relationship between algorithm and application

Speaker verification technologies have many commercial applications, such as direct banking, cellular transactions, credit card operations and E-Commerce. Voice based verification can answer the need for a secure, friendly and cost effective authentication tool required by the finance, commerce and Telecommunication markets. Introducing an operational large-scale system to the market requires much more than a good algorithm. Several design issues should be considered, such as: How to retrieve the audio from the telephony network? What is the optimal way to store and maintain the voice signatures? How to receive claimed identity? Is log likelihood a meaningful score? The development process opens a wide range of subjects for algorithmic research. Among them are: time evolution of speaker models, decision mechanisms, effective scoring, and new ways for constructing world models. Algorithmic research and system development cannot be done independently. Continuous joint work is necessary in order to have successful operational systems, which will make speaker verification the natural authentication means in remote services and transactions. The paper reviews the inter-relationship between algorithmic research and system development based on the experience from the speaker verification product of Persay Ltd. We describe the main problems during the system design process, and discuss the alternatives for solution. A list of research problems, derived from the implementation process is presented.