Abstract: This disclosure relates to a method and an apparatus for training a text classification model. The method may include determining a semantic representation of the training sample using the text classification model and determining a predicted classification result of the training sample based on the semantic representation. The method may further include generating an adversarial sample corresponding to the training sample based on the training sample and perturbation information and determining a semantic representation of the adversarial sample corresponding to the training sample using the text classification model.

Abstract: Database systems and methods are provided for assigning structural metadata to records and creating automations using the structural metadata. One method of assigning structural metadata to a record associated with a conversation involves obtaining a plurality of utterances associated with the conversation, identifying, from among the plurality of utterances, a representative utterance for semantic content of the conversation, assigning the conversation to a group of semantically similar conversations based on the representative utterance, and automatically updating the record associated with the conversation at a database system to include metadata identifying the group of semantically similar conversations.

Abstract: An apparatus for encoding a multi-channel signal having at least two channels, has: a downmixer for calculating a downmix signal from the multi-channel signal; a parameter calculator for calculating a side gain from a first channel of the at least two channels and a second channel of the at least two channels and for calculating a residual gain from the first channel and the second channel; and an output interface for generating an output signal, the output signal having information on the downmix signal, and on the side gain and the residual gain.

Abstract: Implementations are directed to determining an audio delay, of a computing device, by causing an audio data stream to be transmitted to the computing device via a wireless communication channel. The computing device causes audio output generated using the audio data stream to be rendered via speaker(s). The rendered audio output is captured via microphone(s), and the audio delay determined by comparing the captured audio output with the audio data stream. A delay audio segment can be appended to an additional audio data stream transmitted to the computing device, where the length of the delay audio segment is determined using the audio delay. A noise reduction technique can additionally or alternatively be adapted based on the audio delay. Implementations are additionally or alternatively directed to determining if an audio data stream transmitted to a computing device for rendering through speaker(s) driven by the computing device—is actually being rendered.

Abstract: A method for processing an audio signal in accordance with a room impulse response is described. The audio signal is processed with an early part of the room impulse response separate from a late reverberation of the room impulse response, wherein the processing of the late reverberation has generating a scaled reverberated signal, the scaling being dependent on the audio signal. The processed early part of the audio signal and the scaled reverberated signal are combined.

Abstract: A data processing system receives a natural language communication including an ordered sequence of a plurality of word spellings of a natural human language. A processor of the data processing system parses the plurality of word spelling of the natural language communication utilizing constraint-based parsing to identify a plurality of satisfied parsing constraints. The processor performs semantic analysis based on the plurality of satisfied parsing constraints. Performing semantic analysis includes obtaining at least mid-level comprehension of the natural language communication by identifying in the natural language communication utilizing constraints at least one of the following set: a clausal structure within the natural language communication, a sentence structure of a sentence in the natural language communication, an implied topic of the natural language communication, and a classical linguistic role in the natural language communication.

Abstract: A stereo signal processing method includes performing delay estimation on a stereo signal of a current frame to determine an inter-channel time difference of the current frame, identifying a sign of the inter-channel time difference of the current frame is different from a sign of an inter-channel time difference of a previous frame of the current frame, performing delay alignment processing on the first-channel signal of the current frame based on the inter-channel time difference of the current frame, and performing delay alignment processing on the second-channel signal of the current frame based on the inter-channel time difference of the previous frame.

Abstract: One example may include a method that includes initiating an audio recording to capture audio data, comparing the audio data received from a microphone of a mobile device to an audio data range, determining whether the audio data is above an optimal level based on a result of the comparison, and queuing the audio data in an audio data queue when the audio data is above the optimal level.

Abstract: A method for event extraction according to the disclosure includes: processing an object text using a preset extraction model to determine event information of the object text; wherein the event information includes an event element, and an event type and a role corresponding to the event element; and the extraction model includes a classification layer and an output layer; the classification layer is configured to determine a token attribute of a token in the object text; the token attribute includes whether the token is a start token of the event element of any event type and any role, and whether the token is an end token of the event element of any event type and any role; and the output layer is configured to determine the event element according to the token attribute of the token, and determine the event type and the role corresponding to the event element.

Abstract: Example techniques involve noise-robust acoustic echo cancellation. An example implementation may involve causing one or more speakers of the playback device to play back audio content and while the audio content is playing back, capturing, via the one or more microphones, audio within an acoustic environment that includes the audio playback. The example implementation may involve determining measured and reference signals in the STFT domain. During each nth iteration of an acoustic echo canceller (AEC): the implementation may involve determining a frame of an output signal by generating a frame of a model signal by passing a frame of the reference signal through an instance of an adaptive filter and then redacting the nth frame of the model signal from an nth frame of the measured signal. The implementation may further involve determining an instance of the adaptive filter for a next iteration of the AEC.

Abstract: A recording system for a motor vehicle includes a communication interface receiving an audio feed from a telephone. A voice recognition module is communicatively coupled to the communication interface and responds to a command from a human user by converting at least a portion of the audio feed from the telephone into text data, and storing the text data in a memory device.

Abstract: The disclosed systems and methods provide a framework for a proactive prediction of the toxic propensity of an article. Prior to the publication and/or reception of comments to online content, the disclosed framework determines the toxic propensity of the content's context and/or specific words, sentences, sentiments, tone or other messages receivable from consumption of the content. Thus, disclosed framework performs proactive forecasting of the content's toxicity propensity”, which quantifies how likely the content is prone to incur or attract toxic comments. The framework can function and/or be configured to operate in a manner that can perform specifically adherent moderation actions that correspond to the content and control how the content can be interacted with, based on the toxic propensity determination, prior to the content's publication in an effort to thwart, prevent or stop toxic environments surrounding or stemming from the content from coming into existence.

Abstract: A method for decoding an encoded audio bitstream is disclosed. The method includes receiving the encoded audio bitstream and decoding the audio data to generate a decoded lowband audio signal. The method further includes extracting high frequency reconstruction metadata and filtering the decoded lowband audio signal with an analysis filterbank to generate a filtered lowband audio signal. The method also includes extracting a flag indicating whether either spectral translation or harmonic transposition is to be performed on the audio data and regenerating a highband portion of the audio signal using the filtered lowband audio signal and the high frequency reconstruction metadata in accordance with the flag. The high frequency regeneration is performed as a post-processing operation with a delay of 3010 samples per audio channel.

Abstract: Diffuse or spatially large audio objects may be identified for special processing. A decorrelation process may be performed on audio signals corresponding to the large audio objects to produce decorrelated large audio object audio signals. These decorrelated large audio object audio signals may be associated with object locations, which may be stationary or time-varying locations. For example, the decorrelated large audio object audio signals may be rendered to virtual or actual speaker locations. The output of such a rendering process may be input to a scene simplification process. The decorrelation, associating and/or scene simplification processes may be performed prior to a process of encoding the audio data.

Abstract: This disclosure describes a wide band array that uses wideband beamforming with interference cancellation at multiple independent frequencies and spatial locations and main lobe steering at multiple independent frequencies and spatial locations. One embodiment uses one through N analysis filter bands 1410 coupled to one through N narrowband beamformers 1414 with the output processed through one through N synthesis bands and summed together to produce the full spectrum output signal 1426. Another embodiment uses one through M sensors with Discrete Fourier Transforms (DFT) and one through N frequency bins 1412 coupled to one through N narrowband beamformers 1414 processed through an Inverse DFT to produce the full spectrum output signal 1428. Another embodiment uses one through N sensor subarrays and one through N frequency bands 1406 coupled to one through N narrowband beamformers 1414 summed together to produce the full spectrum output signal 1422.

Abstract: Implementations are directed to determining an audio delay, of a computing device, by causing an audio data stream to be transmitted to the computing device via a wireless communication channel. The computing device causes audio output generated using the audio data stream to be rendered via speaker(s). The rendered audio output is captured via microphone(s), and the audio delay determined by comparing the captured audio output with the audio data stream. A delay audio segment can be appended to an additional audio data stream transmitted to the computing device, where the length of the delay audio segment is determined using the audio delay. A noise reduction technique can additionally or alternatively be adapted based on the audio delay. Implementations are additionally or alternatively directed to determining if an audio data stream transmitted to a computing device for rendering through speaker(s) driven by the computing device—is actually being rendered.

Abstract: An apparatus for estimating an inter-channel time difference between a first channel signal and a second channel signal, includes a signal analyzer for estimating a signal characteristic of the first channel signal or the second channel signal or both signals or a signal derived from the first channel signal or the second channel signal; a calculator for calculating a cross-correlation spectrum for a time block from the first channel signal in the time block and the second channel signal in the time block; a weighter for weighting a smoothed or non-smoothed cross-correlation spectrum to obtain a weighted cross correlation spectrum using a first weighting procedure or using a second weighting procedure depending on a signal characteristic estimated by the signal analyzer, wherein the first weighting procedure is different from the second weighting procedure; and a processor for processing the weighted cross-correlation spectrum to obtain the inter-channel time difference.

Abstract: Techniques for configuring device-specific skills as top-level skills are described. When a system receives a user input the system performs NLU processing to determine an intent of the user input. In some instances, the system may identify a device-specific skill associated with the device interacted with by the user. At least partially in parallel to performing NLU processing to determine the intent of the user input, the system may also perform NLU processing to determine a likelihood that the user input corresponds to an intent actionable by the device-specific skill. Once the system has finished NLU processing, the system may implement one or more prioritization rules to determine whether the user input should be sent to the device-specific skill or another skill of the system.

Abstract: Techniques for filtering the output of supplemental content are described. When a supplemental output system (e.g., a supplemental content system or notification system) receives supplemental content for output, the supplemental output system sends a user identifier (of the recipient user) and the supplemental content to separately implemented filtering component. The filtering component uses a machine learning (ML) model to determine a topic of the supplemental content. The filtering component determines whether the supplemental content should not be output based on the ML model-determined topic, one or more guardrail policies of the supplemental output system, and user frustration data regarding previously output supplemental content. Use of the ML model to determine the topic prevents a content publisher from surreptitiously associating supplemental content with a specific topic in an effort to bypass topic-based output guardrails.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing a machine learning task on a tuple of respective input sequences to generate an output. In one aspect, one of the systems includes a neural network comprising a plurality of encoder neural networks and a head neural network, each encoder neural network configured to: receive a respective input sequence from the tuple; process the respective input sequence using one or more encoder network layers to generate an encoded representation comprising a sequence of tokens; and process each of some or all of the tokens in the sequence of tokens using a projection layer to generate a lower-dimensional representation, and the head neural network configured to: receive lower-dimensional representations of a respective proper subset of the sequence of tokens generated by the encoder neural network; and process the lower-dimensional representations to generate the output.