CTT-bank: A speech controlled telephone banking system - an initial evaluation
Håkan Melin, Anna Sandell, Magnus Ihse · 2007
The CTT-bank demonstration system has been created at the Centre for Speech Technology (CTT) at KTH with the purpose of displaying and evaluating speech technology in telephone banking. When calling the system, users say their name and a short digit sequence of their own choice to identify themselves. Speaker verification with a prompted four-digit password is then used to authenticate identity claims. Dialogues are finite state-based. The system asks for a single piece of information at a time, such as an account name or a money amount, and users may respond with keywords optionally embedded in carrier phrases. Three banking services have been implemented: account balance, list of recent transactions and transfer between accounts. Help is always available upon request and users may interrupt the dialogue at any time to go back to the starting point of a service loop. A small user trial with 24 subjects has been performed. Subjects enrolled in the bank and called the service up to seven times during a two-week period. A questionnaire was used to collect user opinions, and speech data was recorded and used for the evaluation of speech recognition and speaker verification components in the system. The average word error rate for speech recognition of utterances with four to seven digits was 4.6%. Concept error rates for recognition of service commands, account names, money amounts and answers to yes/no questions were 12% on average for utterances where all spoken words were represented in the lexicon. The false reject rate for the speaker verification system was 4.7% with 30 seconds of enrolment data and up to three attempts at the 2-second test utterance. Impostor attempts were not included in the trial and therefore no direct measurements of false accept rates can be made. Previous work with user trials on speaker verification in telephone banking contexts has also been reviewed. Based on additional results in those trials, it is estimated that speaker verification error rates around 1% is achievable in a telephone banking application. Combined with other means of authentication this offers a large increase in security. User opinions about the trial were positive, and the overall conclusion is that speech technology is useful in telephone banking.