International Business Machines Corporation

Nuance Communications, Inc.

Marvell International Limited

Intel Corporation

Dialogic Corporation

Hagai Aronowitz

Ron Hoory

Orith Toledo-Ronen

David Nahamoo

Zvi Kons

Jason William Pelecanos

Itzhack Goldberg

Amir Geva

Amit Aides

Ian M Molloy

Lawrence H Koved

Adoram Erell

Bo Zhang

Shay Ben-David

Slava Shechtman

David Dov

An example system includes a processor to receive a data set and similarity scores. The processor is to execute an eigen response analysis on eigenvectors calculated for a similarity matrix generated based on the similarity scores for the data set. The processor is to output an estimated number of clusters in the data set based on the eigen response analysis.

Calculating numbers of clusters in data sets using eigen response analysis

There is provided a computer implemented method of presenting color coded text generated from an audio track of a video, the color coding denoting respective speakers, comprising: receiving the audio track of the video divided into a plurality of audio-segments each representing speech spoken by a respective speaker of a plurality of speakers, for each audio-segment of the plurality of audio-segments: receiving a text representation of the audio-segment, extracting a feature vector from the audio-segment, mapping the feature vector to a color space, coloring the text representation according to the color space, and presenting the colored text representation in association with a video-segment corresponding to the audio-segment.

Systems and methods for automated generation of subtitles

Aspects of the present disclosure relate to enhancing biometric security. A biometric template and a biometric sample can be compared using a first security technique having a first ECC framework with a first number of codewords. A determination can be made that the biometric template and biometric sample do not match using the first security technique. The biometric template and the biometric sample can then be compared using a second security technique having a second ECC framework with a second number of codewords, the second number of codewords less than the first number of codewords. A determination can then be made that the biometric template and the biometric sample match, and a user associated with the biometric sample can be verified.

Biometric data protection

Embodiments of the present systems and methods may provide techniques for handling acoustic-lips timing mismatch in audiovisual processing. In embodiments, the context-dependent time shift between the audio and visual streams may be explicitly modeled using an attention mechanism. For example, in an embodiment, a computer-implemented method for determining a context-dependent time shift of audio and video features in an audiovisual stream or file may comprise receiving audio information and video information of the audiovisual stream or file, processing the audio information and video information separately to generate a new representation of the audio information, including information relating to features of the audio information, and a new representation of the video information, including information relating to features of the video information, and mapping features of the audio information and features of the video information using an attention mechanism to identify synchronized pairs of audio and video features.

Attention mechanism for coping with acoustic-lips timing mismatch in audiovisual processing

Methods, computing systems and computer program products implement embodiments of the present invention that include defining a verification string including a sequence of verification characters and a delimiter character between each sequential pair of the verification characters, the delimiter character being different from the verification characters. The verification string to a user, and upon receiving, from the user, a series of verification vocal inputs in response to presenting the verification string, a set of verification features from each of the verification vocal inputs are computed so as to generate sets of verification features. A one-to-one correspondence is established between each of the verification vocal inputs and each of the verification characters, and the user is authenticated based on the verification vocal inputs and their corresponding sets of verification features.

User authentication using prompted text

Methods, computing systems and computer program products implement embodiments of the present invention that include receiving, at a first time, first video and first audio signals generated in response to a user uttering a passphrase, and receiving, at a second time subsequent to the first time, second video and second audio signals generated in response the user uttering the passphrase. Upon computing an audio temporal alignment between the first and the second audio signals and computing a video temporal alignment between the first and the second video signal, the user can be authenticated by comparing the audio temporal alignment to the video temporal alignment.

User authentication using audiovisual synchrony detection

A method comprising using at least one hardware processor for: providing a set of development supervectors representing features of biometric samples of multiple subjects, the biometric samples being of at least a first and a second different biometric modalities; providing at least a first and a second enrollment supervectors representing features of at least a first and a second enrollment biometric samples of a target subject correspondingly, wherein the at least first and second enrollment samples are of the at least first and the second different biometric modalities correspondingly; providing at least a first and a second verification supervectors representing features of at least a first and a second verification biometric samples of the target subject correspondingly, wherein the at least first and second verification samples are of the at least first and second different biometric modalities correspondingly; concatenating the development supervectors to a set of development generic supervector, the at least first and second enrollment supervectors to a single enrollment generic supervector and the at least first and second verification supervectors to a single verification generic supervector; and verifying an identity of the target subject based on a fused score calculated for the verification generic supervector, wherein the fused score is calculated based on the enrollment generic supervector and the set of development generic supervectors.

Biometric authentication

Dynamic multi-factor authentication challenge selection is provided. A risk associated with an operation that requires authentication of a user of a client device is determined. A plurality of authentication methods is identified. Each respective authentication method associated with a level of security offsetting the risk and a computing cost associated with a respective authentication method. One or more authentication methods are selected from the plurality of authentication methods according to the risk and to minimize the computing cost associated with authenticating the operation.

Dynamic multi-factor authentication challenge generation

A method of detecting an occurrence of splicing in a speech signal includes comparing one or more discontinuities in the test speech signal to one or more reference speech signals corresponding to the test speech signal. The method may further include calculating a frame-based spectral-like representation Sof the speech signal, and calculating a frame-based spectral-like representation Sof a reference speech signal corresponding to the speech signal. The method further includes aligning Sand Sin time and frequency, calculating a distance function associated with aligned Sand S, and evaluating the distance function to determine a score. The method also includes comparing the score to a threshold to detect if splicing occurs in the speech signal.

Method and apparatus for detecting splicing attacks on a speaker verification system

According to some embodiments of the present invention there is provided a computerized method for speech processing using a Gaussian Mixture Model. The method comprises the action of receiving by hardware processor(s) two or more covariance values representing relationships between distributions of speech coefficient values that represent two or more audible input speech signals recorded by a microphone. The method comprises the action of computing two or more eigenvectors and eignevalues using a principle component analysis of the covariance values and transforming the speech coefficient values using the eigenvectors and computing two or more second covariance values from the transformed speech coefficient values. The method comprises the action of modifying some of the second covariance values according to the eignevalues, the covariance values, and two or more indices of the speech coefficient values. The second covariance values to the speech processor comprising the Gaussian Mixture Model.

Covariance matrix estimation with structural-based priors for speech processing

Systems and methods for maintaining speaker recognition performance are provided. A method for maintaining speaker recognition performance, comprises training a plurality of models respectively corresponding to speaker recognition scores from a plurality of speakers over a plurality of sessions, and using the plurality of models to conclude whether a speaker seeking access to an environment is a non-ideal target speaker or a non-ideal non-target speaker. Using the plurality of models to conclude comprises calculating a first probability that the speaker seeking access is the non-ideal target speaker, calculating a second probability that the speaker seeking access is the non-ideal non-target speaker, and determining whether the first probability, the second probability or a sum of the first probability and the second probability is above a probability threshold.

Detection of target and non-target users using multi-session information

A method for inter-dataset variability compensation, the method comprising using at least one hardware processor for: receiving a heterogeneous development dataset comprising multiple samples and metadata associated with at least some of the multiple samples; dividing the multiple samples into multiple homogenous subsets, based on the metadata; averaging high-level features of each of the multiple homogenous subsets, to produce multiple central high-level features for the multiple homogenous subsets, respectively; computing an inter-dataset variability subspace spanned by the multiple central high-level features; removing the inter-dataset variability subspace from the high-level features of the multiple homogenous subsets, to produce denoised samples; and training a machine learning system using the denoised speech samples.

Dataset shift compensation in machine learning

A method for speaker verification is disclosed. The method comprises using at least one hardware processor for: providing a development set comprising multiple voice samples of multiple speakers uttering a predefined development text, prompting a test text to a target speaker, wherein the test text is different from the development text, and recording a test sample of the target speaker uttering the test text, synthesizing a set of artificial voice samples based on the multiple voice samples, wherein each of the artificial voice samples simulates a different speaker of the multiple speakers uttering the test text, and verifying an identity of the target speaker based on the set of artificial voice samples and on the test sample of the target speaker.

Artificial utterances for speaker verification

Computer systems employing speaker verification as a security approach to prevent un-authorized access by intruders may be tricked by a synthetic speech with voice characteristics similar to those of an authorized user of the computer system. According to at least one example embodiment, a method and corresponding apparatus for detecting a synthetic speech signal include extracting a plurality of speech features from multiple segments of the speech signal; analyzing the plurality of speech features to determine whether the plurality of speech features exhibit periodic variation behavior; and determining whether the speech signal is a synthetic speech signal or a natural speech signal based on whether or not a periodic variation behavior of the plurality of speech features is detected. The embodiments of synthetic speech detection result in security enhancement of the computer system employing speaker verification.

Growing community of inventors

Petah-Tikva, Israel

Hagai Aronowitz