Systems, devices, and methods are provided for multi-stem volume equalization, wherein the volume levels of each stem may be adjusted non-uniformly. Audio may be diarized into a plurality of stems, including background noise separate. Mean and variance of the volume levels of the stems may be computed. Each audio stem may be automatically adjusted based on a stem-specific preference that a user may specify. View may adjust actor volume relative to the mean/variance that maintains a relative difference in volume levels between stems.

Techniques for entity and relationship detect from unstructured text as a service are described. A service may receive a request to identify entities within a provided unstructured text element, and the service may segment and tokenize the unstructured text and send the result to multiple services implementing multiple deep machine learning models trained to identify particular entities. The service may send additional requests to an additional service or services implementing additional deep machine learning models to identify relationships between detected attributes and ones of the detected entities. The outputs from all services can be analyzed and consolidated into a single result that identifies the entities, any attributes of the entities, and confidence scores indicating the confidence in each detected entity.

Played audio frames included in first audio content may be received over one or more networks. The first audio content may further include a replaced audio frame. The first audio content may correspond to video content that includes video of a face of a person as the person utters speech that is captured in the first audio content. Location data may also be received over the one or more networks. The location data may indicate locations of facial features of the face of the person in a video frame of the video content. The video frame may correspond to the replaced audio frame. Audio output may be generated that approximates a portion of the speech corresponding to the replaced audio frame. The audio output may be inserted into a replacement audio frame. Second audio content may be played including the played audio frames and the replacement audio frame.

The popularity of videoconferencing has increased rap idly in recent years. Video conferencing tools may allow multiple people at multiple different locations to interact by receiving both audio and video of one another. This may allow for more personal and detailed interactions between people participating in the videoconference, such as opposed to a mere telephone conference. For example, instead of merely hearing each other’s voices, videoconferencing tools may allow participants to see each other’s facial expressions and movements, thereby more closely simulating actual face-to-face interactions between people in the same room or other physical space. While videoconferencing provides many advantages, the transfer of video over bandwidth constrained networks may also involve several challenges. For example, in certain conditions, existing network band width may be insufficient to support all incoming video feeds, thereby resulting in quality degradation, lost frames, interrupted video feeds, and other problems. In some examples, these problems may occur when available band width temporarily drops, when large numbers of participants join the video conference, or for other reasons.

This paper presents Wojood, a corpus for Arabic nested Named Entity Recognition (NER). Nested entities occur when one entity mention is embedded inside another entity mention. Wojood consists of about 550K Modern Standard Arabic (MSA) and dialect tokens that are manually annotated with 21 entity types including person, organization, location, event and date. More importantly, the corpus is annotated with nested entities instead of the more common flat annotations. The data contains about 75K entities and 22.5% of which are nested. The inter-annotator evaluation of the corpus demonstrated a strong agreement with Cohen’s Kappa of 0.979 and an F1-score of 0.976. To validate our data, we used the corpus to train a nested NER model based on multi-task learning using the pre-trained AraBERT (Arabic BERT). The model achieved an overall micro F1-score of 0.884. Our corpus, the annotation guidelines, the source code and the pre-trained model are publicly available.

Negative medical findings are prevalent in clinical reports, yet discriminating them from positive findings remains a challenging task for in-formation extraction. Most of the existing systems treat this task as a pipeline of two separate tasks, i.e., named entity recognition (NER)and rule-based negation detection. We consider this as a multi-task problem and present a novel end-to-end neural model to jointly extract entities and negations. We extend a standard hierarchical encoder-decoder NER model and first adopt a shared encoder followed by separate decoders for the two tasks. This architecture performs considerably better than the previous rule-based and machine learning-based systems. To overcome the problem of increased parameter size especially for low-resource settings, we propose the Conditional Softmax Shared Decoder architecture which achieves state-of-art results for NER and negation detection on the 2010 i2b2/VA challenge dataset and a proprietary de-identified clinical dataset.