Description:[0009] In this patent, AI-Enabled Voice Biometric Customer Authentication System for Contact Center is proposed to reduce the fraudulent account access and reduce the contact centers operational costs. Caller authentication is the process of authenticating a caller's identity through the phone line. Due to increased call volumes in recent years, the need to identify callers securely and efficiently is emerging in Contact Centers (CC). Knowledge Based Authentication (KBA) has been used for many years to identify callers. Both static and dynamic KBA systems are used and enhanced, despite employing advanced KBA, there are still substantial operating costs, security risks, time-consumption, and annoyance to callers and agents. Due to data breaches, personal information is easily accessed and sold on the dark web hence KBA is not a safe method of authentication. Elderly, and uneducated people still find difficulty in KBA process and most vulnerable to fraudulent account access incidents. This invention relates to a system and method of authenticating customers automatically employing the speech biometric technology. Behavioural biometrics are built on a person's unique behavioral pattern, which is difficult to replicate, such as voice. Voice biometrics can analyse a wide range of speech traits, such as behavioural and physical. The combination of all these factors produces a unique voice pattern called a “voiceprint”.
[0010]A voiceprint is as distinct as a fingerprint. Voice authentication is a secure form of identification since it is more difficult to forge. Since it can be done remotely hence more practical than other biometric methods such as iris scans. This invention can be applied in a wide range of scenarios and regions because it is text- and language-independent. Additionally, consumers who are elderly, uneducated, or physically challenged, such as blind persons, can be authenticated more easily by employing this invention. This invention can be an easy way for businesses to cater such customers and serve better. When compared to KBA, this solution provide higher level of security at much lower operational costs. This speech biometric authentication system can be used to authenticate the users/customers in contact centres, smartphone apps, Internet of Things (IoT) devices, and internet services. This development can also be added to multi factor authentication system to enhance the security and reduce the unauthorized account access. Speech biometric offers a genuine answer to the major problems contact centers are facing today. Not only it enhances security, but it frees up agents from the tedious process of question-answer verification and offers a superior level of customer service while using fewer resources and less cost. Decades ago, nearly all transactions were happening in person, therefore it was uncommon to think about customer’s identity verification over voice channels. [0011] But today, with so much activity taking place online or over the phone, connecting to contact center for convenience, the possibility of fraud is surging. Call centres are becoming more aware of this as they handle more complicated activities, such as processing payments and managing enormous amounts of sensitive consumer data. This invention keeps track of all unauthorised access attempts along with information about the attempt account ID, timestamp, and other details. The customer can be informed directly on the phone by a contact center agent to advise customers regarding the unsuccessful unauthorized account access attempts if the number of unauthorised attempts on an account exceeds the predetermined threshold. The same data can be used for reporting to examine the severity, total number of unauthorized accounts attempts, and other analysis objectives. The need of contact center is growing daily in industries including healthcare, banking, financial services, insurance (BFSI), e-commerce, water supply, and telecommunications, among others. Hence reducing the unauthorized account access is utmost priority for contact centers.
[0012]This model requires fewer computing resources. It achieved more than 90% accuracy with 40 speakers. It shows good accuracy even for short duration sample audio files. For some of the speakers, this model is trained using audio files with a duration of 1 to 4 seconds. This model also achieved good accuracy when trained using multiple languages and mixed languages audio files.
[0013] In the new customer enrolment block, Customers are first enrolled into the system taking their text and language independent voice samples. Customers can speak in any language. This invention also supports mixed language and short duration utterances. The voice samples can be in any audio format which is converted to wav format for further process. Voice samples are split into 10 samples per person. In this implementation 10 voice samples per person are taken, also demonstrated that 5 samples per person is also good enough to train the model. These voice samples are processed to convert into MFCC and Delta-MFCC features. Total 40 (20 MFCC and 20 Delta-MFCC) features are extracted from voice samples and trained using Gaussian Mixture Model (GMM) algorithm to create voiceprint for new customers. All the trained GMM speakers’ files/models are saved in a database. Voice samples can be text and language independent to cater any geographic and context.
[0014]A database storage is required in the system to save all the trained GMM models for registered users to authenticate. These are nothing but the voiceprint and contains information of log likelihood. This model requires less computational resources on less cost.
[0015]In the customer authentication block, when registered client calls the Contact center and claim for certain identity, customer’s speech sample is taken on real time basis and transformed into MFCC features, and then compared with the saved voiceprint models for log likelihood matching. Access is granted only if the claimed identity matches the stored voiceprint, or a red flag is raised if claimed identity do not match to the stored voiceprint and stop unauthorized account access. Forging voice is challenging task hence difficult to gain unauthorized access. This model requires fewer computing resources hence cost effective too for small and mid size contact centers and other businesses to apply and leverage more security and reduce day-to-day operational costs.
[0016] The voice recognizer module 201 receives the utterances from customer during new customer enrolment process and later on registered customer authentication process. This system convert audio file to wav format, if needed.
[0017] The feature extraction module 202 takes the wav audio file from module 201. During training phase, voice sample is split into 10 files per person. All the 10 files are processed through feature extraction method to generate MFCC features. Total 40 features are used in this proposed invention implementation comprising 20 MFCC and 20 Delta-MFCC.
[0018] The features system 203 gets these extracted 40 features from previous module and send it to next phase. For new user registration, features are passed to module 204, and for authentication flow features are passed to module 206.
[0019] The speaker modelling module 204 is training phase for new user registration process. All the 40 features for the new user is passed to this phase where GMM modelling is done to generate voiceprint of the user. Voiceprint is similar to fingerprint and unique for every individual. For every user 1 voiceprint is generated after the training. This step calculates log-likelihood for each GMM model of every speaker and stores in a database at next step.
[0020] The speaker models Database is storage system for all the generated voiceprints of registered users. GMM files (voiceprints) are stored in this database.
[0021] The match pattern module 206 is a testing phase. For customer authentication flow, features are diverted from module 203 to 206 for further processing. Data dictionary generated in training phase is used for matching 1: N speaker’s GMM file during testing phase. The speaker with the highest score is chosen.
[0022] Decision phase 207 takes the retrieved customer id from Test phase 206. Logic is written here to match if the claimed identity and retrieved id are same or differ. The decision is calculated in this phase.
[0023] Access grant module 208 approves customers access if the decision from the module 207 is Yes. Customers are connected to contact center after successful authentication for further processing on their accounts.
[0024] Unauthorized module 209 rejects the callers access if the decision from the module 207 is No and raise a red flag which can be later notified to customer for unauthorized account access attempt.
[0025] Unauthorized access trials database 210 is used to store the details of all the unauthorized account access attempts such as tried account ID, timestamp and other details. If the unauthorized attempts on an account exceed set threshold, then customer can be connected directly on the phone line by contact center representative to inform on the unsuccessful unauthorized account access to alert.
[0026] Failed attempts reporting module 211 retrieves data captured in module 210 and can be used to analyse the number of unauthorised accounts attempts and counts and severity. This process can help contact center to operate more securely at lower cost and with less manual intervention.

, Claims:1. AI-Enabled Voice Biometric Customer Contact Center Authentication System comprising;
a. Gaussian Mixture Model (GMM) - Mel Frequency Cepstral Coefficients (MFCC) Voice biometric customer authentication system for contact centers
b. Text independent Voice biometric customer authentication system,
c. Language independent Voice biometric customer authentication system.
d. Unauthorized unsuccessful account attempts recording system to advise customers
e. Unauthorized unsuccessful account attempts recording system to share data to teams for analysis and other objectives.

2. The system of claim 1, wherein GMM-MFCC Voice biometric authentication system using Machine Learning with python codes

3. The system of claim 1, wherein Text Independent Voice biometric using Machine Learning with python codes

4. The system of claim 1, wherein Language Independent Voice biometric using Machine Learning with python codes

5. The system of claim 1, wherein unauthorized unsuccessful account attempts are recorded in the database using python codes

6. The system of claim 1, wherein unauthorized unsuccessful account attempts recording system to share data to teams for analysis and other objectives using with python codes