US 12,738,274 B1Granted

Detecting statistical anomalies in voice interactions

AssigneeCitibank, N.A.

InventorsJames Randolph Myers, Jefferson Okraku

Life of this patent

Estimating the term…

Loading patent…

Abstract

The systems and methods disclosed herein detect indicators of undesirable intent during interactions between users and digital avatars by evaluating emotional signals in user speech. The system generates emotion feature vectors from user input streams that capture characteristics such as pitch variation, speech rate, vocal intensity, spectral features, and so forth. The system uses a two-level evaluation that compares the emotion feature vector against a population baseline derived from historical user interactions to generate a population anomaly score and further compares the emotion feature vector against an allow list database and a deny list database to generate a threat match score based on known protected categories (e.g., threat categories, allowed categories). The system aggregates these scores to produce a risk score, which is evaluated against threshold values to determine a response action (e.g., logging the interaction, injecting clarifying questions, routing to human agents, terminating the interaction).

Claims20

Dependent claims

A system for modeling a physical environment, the system comprising:

at least one hardware processor;
and
at least one non-transitory memory storing instructions, which, when executed by the at least one hardware processor, cause the system to:
receive, during an interaction between a user and a conversational artificial intelligence (AI) avatar, an audio stream associated with the user;
generate, using the audio stream, an emotion feature vector that comprises a value set corresponding to one or more of:
a pitch variation, a speech rate, a vocal intensity, or a spectral characteristic of the audio stream;
perform a first evaluation by comparing the emotion feature vector against a population database to generate a population anomaly score,
wherein the population database comprises a first plurality of historical feature vectors derived from a first historical user interaction set, and
wherein the population anomaly score is generated by determining a distance between the emotion feature vector and one or more cluster centroids derived from the first plurality of historical feature vectors associated with a plurality of other users in the population database;
perform a second evaluation by comparing the emotion feature vector against a protected database to generate a threat match score,
wherein the protected database comprises a second plurality of historical feature vectors derived from a second historical user interaction set different from the first historical user interaction set,
wherein each of the second plurality of historical feature vectors represents one or more protected categories, and
wherein the threat match score is generated by applying an AI model configured to output a likelihood that the emotion feature vector corresponds to the one or more protected categories;
generate a risk score by aggregating the population anomaly score from the first evaluation and the threat match score from the second evaluation;
determine a response action using the risk score by evaluating the risk score against one or more threshold values derived from the population database and the protected database;
and
execute the response action by transmitting an instruction set associated with the response action as an input into the conversational AI avatar.

14 more elements

A non-transitory, computer-readable storage medium comprising instructions thereon, wherein the instructions, when executed by at least one data processor of a system, cause the system to:

access an input stream associated with an interaction between a user and an artificial intelligence (AI)-based avatar;
generate, using the input stream, an emotion feature vector that comprises a value set corresponding to one or attributes of the input stream;
comparing a vector representation of the input stream against a first database to generate a first score corresponding to a population anomaly score,
wherein the first database corresponds to a population database and includes a first historical feature vector set that corresponds to a first historical user interaction set;
comparing the vector representation of the input stream against a second database to generate a second score corresponding to a threat match score,
wherein the second database corresponds to a protected category database and includes a second historical feature vector set that corresponds to a second historical user interaction set, and
wherein the second score represents a likelihood that the emotion feature vector corresponds to one or more protected categories indicated by the second historical user interaction set;
determine a third score corresponding to a risk score based on the first score and the second score;
determine an action using the third score by evaluating the third score against one or more criterion;
and
cause execution of the action by transmitting an instruction set associated with the action as an input into the AI-based avatar.

8 more elements

Dependencies are inferred from the claim text.