Abstract: Techniques for parallel execution component initialization using static dependency analysis are disclosed. In an example method, a computing system receives a set of program code including execution components. The computing system generates a dependency graph of the execution components. The computing system determines, from the dependency graph, a subset of the execution components that do not depend on any other execution component. The computing system configures a parallel process initiator to initialize, in parallel, the subset of the execution components. The computing system, responsive to a first execution component of the first subset of the execution components completing initialization, determines, from the dependency graph, a second execution component from the subset of the execution components that depends on the first execution component and has no other unsatisfied dependencies. The computing system configures the parallel process initiator to initialize the second execution component.

Abstract: Techniques for cross-lingual voice cloning for low-resource languages are provided. In an example method, a computing system receives a first speech sample, the first speech sample characterized by a first voice and spoken in a first language. The computing system receives a first text input in a second language. The computing system generates, by a first ML model, a first encoded representation of the first text input. The computing system generates, by a second ML model, a first spectrogram of the first text input spoken in the first voice in the second language, in which the generating is based on an embedded representation of the first speech sample, in which the first ML model is different from the second ML model. The computing system generates an audio output based on the first spectrogram.

Abstract: A conference gallery view intelligence system determines at least two regions of interest within a conference room based on an input video stream received from a video capture device located within the conference room. An output video stream for rendering within conferencing software is produced for each of the at least two regions of interest. The output video stream for each of the at least two regions of interest is then transmitted to one or more client devices connected to the conferencing software.

Abstract: Techniques for implementing task inference using an AI interface are disclosed. In an example method, a computing system receives, from a first client device, a first chat message. The computing system determines, using a language model, a task based on the first chat message, the task including one or more executable instructions. The computing system outputs, to the first client device, information about the task. The computing system receives, from the first client device, a first indication to perform the task. The computing system outputs one or more commands to cause an execution of the one or more executable instructions of the task and a generation of a task output. The computing system outputs, to the first client device, the task output.

Abstract: Dynamic selection of participants in video conference where a host of a live session can select a list of participants and a rule for selecting participants therefrom, causing video conferencing services to iteratively apply the rule to the list to select participants for active engagement in the live session. The host can select multiple lists to generate combined lists using set operators such as union, intersection, and difference. The host can modify parameters of parameterized rules. A console may be presented by a user interface of a client device of the host, where the host can select and combine lists, and select and modify rules. The console may present a queue of selected participants.

Abstract: Methods and systems provide for a sign language interpreter view within a video communication session. In one embodiment, a method presents a user interface (“UI”) for each of a number of client devices connected to a communication session, each UI including one or more video feeds associated with participants of the communication session. The method receives a request from a first participant to be presented with a sign language interpreter (“SLI”) view within the UI; determines that a second participant is to be designated as a sign language interpreter for the communication session; and presents, within the UI of a client device associated with the first participant, an SLI view including at least a video feed of the second participant and a video feed of a highlighted speaker being interpreted by the second participant.

Abstract: The present disclosure relates to systems and methods for providing assistance to visually impaired and/or hearing-impaired users accessing a video conference. The systems and methods include initiating a virtual meeting hosting a plurality of participants, detecting a presence of at least one impaired participant from the plurality of participants, loading one or more audio visual settings for providing assistance to the at least one impaired participant, and automatically providing adjustments for one or more audio visual elements provided to the at least one impaired participant, based on the one or more audio visual settings.

Abstract: One example method includes joining, by a client application executed by a client device, a virtual conference hosted by a virtual conference provider, a plurality of participants attending the virtual conference; receiving, by the client application, a question associated with the virtual conference; generating a query context based on a real-time transcript of the virtual conference; providing the query context and the question to a trained large language model (“LLM”); and receiving a response from the LLM based on the question and the query context.

Abstract: Techniques for scoped querying using customized artificial intelligence are provided. In an example method, a computing device determines a set of documents about a user of a first client device of a plurality of client devices. The computing device integrates the set of documents into a datastore, the datastore accessible by a machine learning model trained to respond to natural-language queries in the context of the set of documents. The computing device receives, from a second client device of the plurality of client devices, a query about the first user of the first client device and then receives, from the machine learning model, a response to the query based on the set of documents. The computing device outputs the response to the second client device.

Abstract: Techniques for multi-microphone audio alignment are disclosed. An example method involves buffering multiple audio streams from multiple audio input devices for a duration. The method further involves generating multiple sets of transformed buffered audio streams by transforming each buffered audio stream using one or more transformation techniques. The method further involves, for each set of transformed buffered audio streams, determining one or more audio offsets for the corresponding audio stream using the multiple sets of transformed buffered audio streams. The method further involves, for each buffered audio stream, applying the determined one or more audio offsets to the buffered audio stream. The method further involves generating an audio output stream using the multiple buffered audio streams.

Abstract: Systems and methods of acoustic echo cancelation for a multi-microphone conference system are provided. A video conference device receives an audio signal comprising speech associated with a currently speaking participant during a hybrid communication session. The video conference device includes a dedicated microphone and is communicatively coupled to one or more extended microphones related to one or more local client devices associated with one or more local participants. The video conference device determines whether the currently speaking participant is a remote participant or a local participant based on the speech signal. The video conference device activates, based on whether the currently speaking participant is a remote participant or a local participant, one or more microphones of the dedicated microphone and the one or more extended microphones.

Abstract: Techniques for maintaining audio quality during periods of missing data for multi-microphone systems are disclosed. In an example method, a computing system receives multiple audio streams from multiple audio capture devices, each audio stream populating a buffer of multiple buffers including a first buffer, a second buffer, and a reference buffer. The method further involves determining that a first quantity of audio data in the reference buffer exceeds a first predetermined threshold. The method further involves determining that a second quantity of audio data in the first buffer is below a second predetermined threshold. The method further involves applying an audio reconstruction technique to the first buffer to generate additional audio data to add to the first buffer. The method further involves generating an audio output stream using the reference buffer, the first buffer, and the second buffer.

Abstract: Scrolling motion is detected within a video stream to output an indication of a scrolling motion vector for use in encoding a current picture of the video stream. A first line of pixels within a motion region of the current picture is identified. A second line of pixels matching the first line of pixels is identified within a last played picture of the video stream. The scrolling motion vector is determined based on a comparison of lines of pixels nearby the second line of pixels within the last played picture. The indication of the scrolling motion vector is then output for use in encoding the current picture.

Abstract: Techniques for fine-tuning language models using target language vocabulary and parallel data for machine translation are disclosed. In an example method, a computing device receives one or more words in a source language. The computing device translates the one or more words to a target language using a fine-tuned language model. The language model is fine-tuned by accessing a foundation language model pre-trained using source language training data and having a source language vocabulary. A set of tokens is added to the model representing a target language vocabulary. The model is further trained using target language training data. The model is further trained using parallel training data including a number of parallel elements each parallel element including words in the source language and the corresponding words in the target language. The computing device outputs the translated one or more words in the target language.

Abstract: Different visual content simultaneously output for viewing to different people within a physical location (e.g., a conference room) is determined based on viewing angles of those people relative to a display of the physical location. In one approach, viewing angles of conference participants are determined with respect to a three-dimensional glasses-less display within a physical location. First content to output at the three-dimensional glasses-less display for viewing by a first set of the conference participants is determined. Second content to output at the three-dimensional glasses-less display for viewing by a second set of the conference participants is determined. The three-dimensional glasses-less display is then caused to simultaneously output the first content and the second content in a manner that limits visibility of the first content to the first set of the conference participants and visibility of the second content to the second set of the conference participants.

Abstract: A server identifies a condition impacting communication services at a premises housing devices. The server transmits, to multiple devices of multiple respective administrators, a message indicating the condition. The server creates a communication session for addressing the condition. The server adds accounts associated with the multiple devices of the multiple administrators to the communication session.

Abstract: Inter-party communications, such as telephony communications, virtual meeting communications, or messaging communications, are switched between clients while in progress. An inter-party communication may be detected at a first client. A channel may then be opened between the first client and a second client responsive to an operator initiating switching the inter-party communication from the first client to the second client. The inter-party communication may be switched from the first client to the second client using the channel while the inter-party communication remains in progress. The inter-party communication may then be continued at the second client.

Abstract: One example method includes determining, by a computing device, a user present in a room is associated with a virtual space hosted at a video conference provider, wherein the room includes one or more peripheral devices; connecting, by the computing device, to the virtual space; receiving, by the computing device, data from at least one of the one or more peripheral devices or one or more user devices while connected to the virtual space, the one or more user devices associated with the user or one or more other users within the room; transmitting, by the computing device, at least a portion of the data to the virtual space; determining, by the computing device, a state change associated with the room based on the data received from the one or more peripheral devices or one or more user devices; generating, by the computing device, a change detection signal; and transmitting, by the computing device, the change detection signal to the video conference provider to cause the video conference provider to store d

Abstract: Proactive communication service event system includes aggregating phone data indicative of call connection quality for one or more phone devices associated with a customer of a software platform over a telephony network implemented by the software platform. A service event affecting the one or more phone devices is determined based on the aggregated phone data. The aggregated phone data and a representation of the service event is output for display within a graphical user interface. The graphical user interface is rendered at an administrator device of the customer. One or more telephony network resources, corresponding to the service event, of the customer are changed based on an input received via the graphical user interface.

Abstract: A method includes connecting a call from a client device to a destination having an interactive voice response service; transcribing audio from the destination during the call to identify menu options of the interactive voice response service; generating visualizations representing the menu options; and outputting the visualizations to a display associated with the client device. A system includes a telephony system, an automatic speech recognition processing tool, and a visualization output generation tool. The telephony system connects a call from a client device to a destination having an interactive voice response service. The automatic speech recognition processing tool transcribes audio from the destination during the call to identify menu options of the interactive voice response service. The visualization output generation tool generates visualizations representing the menu options. The telephony system outputs the visualizations to a display associated with the client device.