Abstract: Techniques for avoiding wake word self-triggering are provided. In one embodiment, an electronic device can receive an audio-out signal to be output as audio via a speaker of the device and can attempt to recognize a wake word in the audio-out signal using a first recognizer. If the wake word is recognized in the audio-out signal, the electronic device can further determine whether a wake word match is made using a second recognizer with respect to a mic-in audio signal captured via a microphone of the device at approximately the same time that the audio-out signal is output via the speaker. If so, the electronic device can ignore the wake word match made using the second recognizer.

Abstract: Smart listening modes for supporting quasi always-on listening on an electronic device are provided. In one embodiment, the electronic device can determine that a user is likely to utter a voice trigger in order to access the always-on listening functionality of the electronic device. In response to this determination, the electronic device can automatically enable the always-on listening functionality. Similarly, the electronic device can determine that a user is no longer likely to utter the voice trigger in order to access the always-on listening functionality of the electronic device. In response to this second determination, the electronic device can automatically disable the always-on listening functionality.

Abstract: Techniques for implementing dynamic volume adjustment by a virtual assistant are provided. In one embodiment, the virtual assistant can receive a voice query or command from a user, recognize the content of the voice query or command, process the voice query or command based on the recognized content, and determine an auditory response to be output to the user. The virtual assistant can then identify a plurality of criteria for automatically determining an output volume level for the response, where the plurality of criteria including content-based criteria and environment-based criteria, calculate values for the plurality of criteria, and combine the values to determine the output volume level. The virtual assistant can subsequently cause the auditory response to be output to the user at the determined output volume level.

Abstract: Systems, methods, and devices for location-based context driven speech recognition are disclosed. A mobile or stationary computing device can include position locating functionality for determining the particular physical location of the computing device. Once the physical location of the computing device determined, a context related to that particular physical location. The context related to the particular physical location can include information regarding objects or experiences a user might encounter while in that particular physical location. The context can then be used to determine delimited or constrained speech recognition vocabulary subset based on the range of experiences a user might encounter within a particular context.

Abstract: Techniques for implementing face-controlled liveness verification are provided. In one embodiment, a computing device can present, to a user, a sequential series of targets on a graphical user interface (GUI) of the computing device, where each target is a visual element designed to direct the user's attention to a location in the GUI. The computing device can further determine whether the user has successfully hit each target, where the determining comprises tracking movement of a virtual pointer controlled by the user's gaze or face pose and checking whether the user has moved the virtual pointer over each target. If the user has successfully hit each target, the computing device can conclude that the user is a live subject.

Abstract: Techniques for unobtrusively verifying the identity of a user of a computing device are provided. In one embodiment, the computing device can establish one or more verification models for verifying the user's identity, where at least a subset of the one or more verification models is based on enrollment data that is collected in an unobtrusive manner from the user. The computing device can then verify the user's identity using the one or more verification models.

Abstract: Techniques for performing low-fidelity always-on A/V monitoring are provided. In one embodiment, an always-on A/V monitoring system can record audio or video footage of an area of interest on a continuous basis while operating in a low-fidelity recording mode, where the recorded audio or video footage has a quality level that is sufficient to detect one or more events that have meaning to the system or a user, but is insufficient to recognize details with respect to the area of interest that would be considered private to an individual appearing in, or associated with, the recorded audio or video footage.

Abstract: Techniques for implementing voice-based liveness verification are provided. In one embodiment, a computing device can present a series of challenge prompts to a user being authenticated, where each challenge prompt corresponds to a request to utter a liveness passphrase that is randomly selected from a set of liveness passphrases that have been previously enrolled by an enrolled user of the computing device. The computing device can then receive utterances from the user in response to the series of challenge prompts and, if each utterance matches its corresponding enrolled liveness passphrase, can conclude that the user is a live subject.

Abstract: Techniques for implementing unobtrusive training for speaker verification are provided. In one embodiment, an electronic device can receive a plurality of voice samples uttered by one or more users as they interact with a voice command-and-control feature of the electronic device and, for each voice sample, assign the voice sample to one of a plurality of voice type categories. The electronic device can further group the voice samples assigned to each voice type category into one or more user sets, where each user set comprises voice samples likely to have been uttered by a unique user. The electronic device can then, for each user set: (1) generate a voice model, (2) issue, to the unique user, a request to provide an identity or name, and (3) label the voice model with the identity or name provided by the unique user.

Abstract: Techniques for implementing face-controlled liveness verification are provided. In one embodiment, a computing device can present, to a user, a sequential series of targets on a graphical user interface (GUI) of the computing device, where each target is a visual element designed to direct the user's attention to a location in the GUI. The computing device can further determine whether the user has successfully hit each target, where the determining comprises tracking movement of a virtual pointer controlled by the user's gaze or face pose and checking whether the user has moved the virtual pointer over each target. If the user has successfully hit each target, the computing device can conclude that the user is a live subject.

Abstract: Techniques for analyzing long-term audio recordings are provided. In one embodiment, a computing device can record audio captured from an environment of a user on a long-term basis (e.g., on the order of weeks, months, or years). The computing device can store the recorded audio on a local or remote storage device. The computing device can then analyze the recorded audio based one or more predefined rules and can enable one or more actions based on that analysis.

Abstract: Techniques for leveraging a combination of audio-based and vision-based cues for voice command-and-control are provided. In one embodiment, an electronic device can identify one or more audio-based cues in a received audio signal that pertain to a possible utterance of a predefined trigger phrase, and identify one or more vision-based cues in a received video signal that pertain to a possible utterance of the predefined trigger phrase. The electronic device can further determine a degree of synchronization or correspondence between the one or more audio-based cues and the one or more vision-based cues. The electronic device can then conclude, based on the one or more audio-based cues, the one or more vision-based cues, and the degree of synchronization or correspondence, whether the predefined trigger phrase was actually spoken.

Abstract: Techniques for leveraging multiple biometrics for enabling user access to security metadata are provided. In one embodiment, a computing device can receive first and second biometric identifiers from a user. The computing device can further determine, via a multi-biometric authentication system, that the user's identity can be verified using the first biometric identifier, but cannot be, or has not been, verified using the second biometric identifier. In response to this determination, the computing device can provide information to the user for facilitating verification of the user's identity using the second biometric identifier.

Abstract: In one embodiment, a computing device can detect an utterance of a target phrase within an acoustic input signal. The computing device can further determine a first estimate of cumulative signal and noise energy for the detected utterance in the acoustic input signal with respect to a first time period spanning the duration of the detected utterance, and a second estimate of noise energy in the acoustic input signal with respect to a second time period preceding (or following) the first time period. The computing device can then calculate a signal-to-noise ratio (SNR) for the detected utterance based on the first and second estimates and can reject the detected utterance if the SNR is below an SNR threshold.

Abstract: In one embodiment the present invention includes a method comprising receiving an acoustic input signal and processing the acoustic input signal with a plurality of acoustic recognition processes configured to recognize the same target sound. Different acoustic recognition processes start processing different segments of the acoustic input signal at different time points in the acoustic input signal. In one embodiment, initial states in the recognition processes may be configured on each time step.

Abstract: Techniques for performing continuous enrollment for face verification are provided. In one embodiment, a computing device can receive, from a user, an indication that the user wishes to authenticate himself/herself with the computing device via face verification. In response to the indication, the computing device can capture, using a camera, a series of images of the user's face and can authenticate the user by evaluating each of the series of images against a face template for the user, where the user is authenticated based on an N-th image in the series. Once the user has been authenticated, the computing device can select one or more images from the series prior to the N-th image and can add the selected images to the user's face template.

Abstract: Techniques for implementing face-based authentication with situational adaptivity are provided. In one embodiment, a computing device can create an enrollment template for a user, the enrollment template being derived from one or more enrollment images of the user's face and being usable by a face-based authentication system to authenticate the user's identity. The computing device can further determine a first set of metadata associated with the enrollment image(s) and can store the first set of metadata with the enrollment template. At a later time (e.g., an authentication event), the computing device can capture an input image of the user's face, determine a second set of metadata associated with the input image, and calculate a computational distance between the input image and the enrollment template, the calculating taking into account a degree of difference between the first and second sets of metadata. Finally, the user can be authenticated based on the distance.

Abstract: In one embodiment, a computer system stores speech data for a plurality of speakers, where the speech data includes a plurality of feature vectors and, for each feature vector, an associated sub-phonetic class. The computer system then builds, based on the speech data, an artificial neural network (ANN) for modeling speech of a target speaker in the plurality of speakers, where the ANN is configured to discriminate between instances of sub-phonetic classes uttered by the target speaker and instances of sub-phonetic classes uttered by other speakers in the plurality of speakers.

Abstract: In one embodiment, a computing device can detect an utterance of a target phrase within an acoustic input signal. The computing device can further determine a first estimate of cumulative signal and noise energy for the detected utterance in the acoustic input signal with respect to a first time period spanning the duration of the detected utterance, and a second estimate of noise energy in the acoustic input signal with respect to a second time period preceding (or following) the first time period. The computing device can then calculate a signal-to-noise ratio (SNR) for the detected utterance based on the first and second estimates and can reject the detected utterance if the SNR is below an SNR threshold.

Abstract: In one embodiment, a method includes receiving an acoustic input signal at a speech recognizer. A user is identified that is speaking based on the acoustic input signal. The method then determines speaker-specific information previously stored for the user and a set of responses based on the recognized acoustic input signal and the speaker-specific information for the user. It is determined if the response should be output and the response is outputted if it is determined the response should be output.