Microsoft Technology Licensing, LLC

Dropbox, Inc.

Ranjitha Gurunath Kulkarni

Jongmin Baek

Imed Zitouni

Ziad Al Bawab

Ermo Wei

Ahmed Moustafa El Kholy

Noha Alon

Xingyu Xiang

James Johnson

Shuangyu Chang

Serge-Eric Tremblay

Anirudh Koul

Rajkumar Janakiraman

Meryem Pinar Donmez Ediz

Nitin Kamra

The present disclosure relates to systems, non-transitory computer-readable media, and methods for generating responses to prompts by utilizing a context engine and a logic engine. In one or more embodiments, the disclosed systems can determine a prompt received from a client device involves one or more logical problems with a prompt classification model. Based on identifying the one or more logical problems, the disclosed systems can generate a logic code segment by processing the prompt with one or more large language models within a context engine. The disclosed systems can generate a logic result for the prompt by processing the logic code segment with a logic engine that solves one or more logical problems within the prompt according to the structure of the logic code segment. The disclosed systems can generate a response to the prompt based, at least in part, on the logic result.

Generating responses using a context engine coupled with a logic engine and time phrase resolution

The present disclosure relates to systems, methods, and non-transitory computer-readable media that can leverage a natural language model to determine a most probable candidate sequence of tokens and thereby generate a predicted user activity. In particular, the disclosed systems can tokenize activity event vectors to generate a series of sequential tokens that correspond to recent user activity of one or more user accounts. In addition, the disclosed systems can, for each candidate (e.g., hypothetical) user activity, augment the series of sequential tokens to include a corresponding token. Based on respective probability scores for each of the augmented series of sequential tokens, the disclosed systems can identify as the predicted user activity, a candidate user activity corresponding to one of the augmented series of sequential tokens associated with a highest probability score. Based on the predicted user activity, the disclosed systems can surface one or more suggestions to a client device.

Utilizing a natural language model to determine a predicted activity event based on a series of sequential tokens

Automatic speech recognition systems can benefit from cues in user voice such as hyperarticulation. Traditional approaches typically attempt to define and detect an absolute state of hyperarticulation, which is very difficult, especially on short voice queries. This disclosure provides for an approach for hyperarticulation detection using pair-wise comparisons and on a real-world speech recognition system. The disclosed approach uses delta features extracted from a pair of repetitive user utterances. The improvements provided by the disclosed systems and methods include improvements in word error rate by using hyperarticulation information as a feature in a second pass N-best hypotheses rescoring setup.

Hyperarticulation detection in repetitive voice queries using pairwise comparison for improved speech recognition

Disclosed in various examples are methods, systems, and machine-readable mediums for providing improved computer implemented speech recognition by detecting and correcting speech recognition errors during a speech session. The system recognizes repeated speech commands from a user in a speech session that are similar or identical to each other. To correct these repeated errors, the system creates a customized language model that is then utilized by the language modeler to produce a refined prediction of the meaning of the repeated speech commands. The custom language model may comprise clusters of similar past predictions of speech commands from the speech session of the user.

Correction of speech recognition on repetitive queries

One or more systems and/or techniques are provided for automatic closed captioning for media content. In an example, real-time content, occurring within a threshold timespan of a broadcast of media content (e.g., social network posts occurring during and an hour before a live broadcast of an interview), may be accessed. A list of named entities, occurring within the social network data, may be generated (e.g., Interviewer Jon, Interviewee Kathy, Husband Dave, Son Jack, etc.). A ranked list of named entities may be created based upon trending named entities within the list of named entities (e.g., a named entity may be ranked higher based upon a more frequent occurrence within the social network posts). A dynamic grammar (e.g., library, etc.) may be built based upon the ranked list of named entities. Speech recognition may be performed upon the broadcast of media content utilizing the dynamic grammar to create closed caption text.

Automated closed captioning using temporal data

Growing community of inventors

Sunnyvale, CA, United States of America

Ranjitha Gurunath Kulkarni