Back to results
US 10,593,336 B2Granted
Machine learning for authenticating voice
AssigneesAccenture Global Solutions Limited, Misram LLC
InventorsConstantine T. Boyadjiev, Rajarathnam Chandramouli, Koduvayur Subbalakshmi, Zongru Shao
Life of this patent
Estimating the term…
Abstract
A machine learning multi-dimensional acoustic feature vector authentication system, according to an example of the present disclosure, builds and trains multiple multi-dimensional acoustic feature vector machine learning classifiers to determine a probability of spoofing of a voice. The system may extract an acoustic feature from a voice sample of a user. The system may convert the acoustic feature into multi-dimensional acoustic feature vectors and apply the multi-dimensional acoustic feature vectors to the multi-dimensional acoustic feature vector machine learning classifiers to detect spoofing and determine whether to authenticate a user.
Claims20
A machine learning multi-dimensional acoustic feature vector authentication system comprising:
- at least one processor to execute machine readable instructions stored on at least one non-transitory computer readable medium;
- at least one data storage to store a plurality of multi-dimensional acoustic feature vector machine learning classifiers,
- wherein the plurality of multi-dimensional acoustic feature vector machine learning classifiers comprise convolutional neural networks trained to identify multi-dimensional acoustic feature vectors;
- wherein the at least one processor is to execute machine readable instructions to:
- extract at least one acoustic feature from a voice sample of a user;
- convert the acoustic feature into a plurality of multi-dimensional acoustic feature vectors;
- apply each multi-dimensional acoustic feature vector in the plurality of multi-dimensional acoustic feature vectors to a corresponding multi-dimensional acoustic feature vector machine learning classifier from the plurality of multi-dimensional acoustic feature vector machine learning classifiers;
- determine a probability of spoofing for each multi-dimensional acoustic feature vector from an output of the corresponding multi-dimensional acoustic feature vector machine learning classifier;
- determine an overall probability of spoofing for the voice sample, based on the probability of spoofing for each multi-dimensional acoustic feature vector;
- determine whether the overall probability of spoofing is within a certain threshold;
- and
- in response to a determination that the overall probability of spoofing is within the certain threshold, authenticate the user.
9 more elements
A machine learning multi-dimensional acoustic feature vector authentication system comprising:
- at least one processor to execute machine readable instructions stored on at least one non-transitory computer readable medium;
- at least one data storage to store a plurality of multi-dimensional acoustic feature vector machine learning classifiers trained to identify multi-dimensional acoustic feature vectors;
- wherein the at least one processor is to execute the machine readable instructions to:
- extract at least one acoustic feature from a voice sample of a user;
- convert the at least one acoustic feature into a plurality of multi-dimensional acoustic feature vectors using a Short-Time Fourier Transformation, a Mel-Frequency Cepstral Coefficient Transformation, a Tempo-gram Feature and a Tonnetz space Geometric representation algorithm;
- apply each multi-dimensional acoustic feature vector in the plurality of multi-dimensional acoustic feature vectors to a corresponding multi-dimensional acoustic feature vector machine learning classifier from the plurality of multi-dimensional acoustic feature vector machine learning classifiers, wherein the plurality of multi-dimensional acoustic feature vector machine learning classifiers is trained to detect multi-dimensional acoustic feature vectors that are converted using the Short-Time Fourier Transformation, the Mel-Frequency Cepstral Coefficient Transformation, the Tempo-gram Feature and the Tonnetz space geometric representation;
- determine a probability of spoofing for each multi-dimensional acoustic feature vector from an output of the corresponding multi-dimensional acoustic feature vector machine learning classifier;
- determine a confidence interval for the probability of spoofing for each multi-dimensional acoustic feature vector;
- determine a weight for the probability of spoofing for each multi-dimensional acoustic feature vector based on the confidence interval;
- determine an overall probability of spoofing for the voice sample, based on a weighted combination of the probability of spoofing for each multi-dimensional acoustic feature vector in the multi-dimensional acoustic feature vectors, wherein the weighted combination is based on the weight for each multi-dimensional acoustic feature vector;
- determine whether the overall probability of spoofing is within a certain threshold;
- and
- in response to a determination that the overall probability of spoofing is within the certain threshold, authenticate the user.
10 more elements
A method of training a machine learning multi-dimensional acoustic feature vector authentication system, the method comprising:
- receiving test voice samples that correspond to a plurality of test group users;
- determining known spoofed voice samples that correspond to the test voice samples using a first known spoofing method;
- determining a training dataset that includes the test voice samples and the corresponding known spoofed voice samples;
- determining multi-dimensional acoustic feature vectors for the training dataset based on a first multi-dimensional acoustic feature vector extraction algorithm;
- applying each multi-dimensional acoustic feature vector for the training dataset to a multi-dimensional acoustic feature vector machine learning classifier;
- determining a first multi-dimensional acoustic feature vector model for the training dataset;
- determining an unknown spoofed voice sample that corresponds to a test voice sample of the test voice samples using an unknown spoofing method;
- determining an unknown multi-dimensional acoustic feature vector of the unknown spoofed voice sample;
- applying the unknown multi-dimensional acoustic feature vector to the multi-dimensional acoustic feature vector machine learning classifier using the first multi-dimensional acoustic feature vector model;
- determining a probability of spoofing for each multi-dimensional acoustic feature vector from an output of the multi-dimensional acoustic feature vector machine learning classifier;
- and
- determining a confidence interval for the probability of spoofing for the first multi-dimensional acoustic feature vector model.
9 more elements
Dependencies are inferred from the claim text.