Microsoft Technology Licensing, LLC

Marchex, Inc.

Jason S Flaks

Ziad Ismail

Chris Kolbegger

Avi Bar-Zeev

Bryan Rieger

Alex Aben-Athar Kipman

Jeffrey Neil Margolis

Christian Klein

John Allen Tardif

Ivan Jelev Tashev

Andrew John Fuller

Jai Srinivasan

Ali M Vassigh

Matthew Berk

Mark James Encarnacion

Govindaraj Ramanathan

Darren Spehr

Chris Miles

Shane Walker

Iroro Orife

Tim Graber

Alejandro Acero

Thomas M Soemo

John Michael Thornton

Robert Norman Heitkamp

Jordan L K Schwartz

Michael I Torres

Rudy Jacobus Poot

Joseph Andrew Bono

Duncan James McKay

Ryan A Kabir

Vanessa Larco

Shahram Izadi

Daniel F Plastina

Tommer Leyvand

Mark James Finocchio

Michael Isard

Robert L Crocco, Jr

Alexandru Octavian Balan

David Kim

Michael Baseflug

Wei Guo

David G Molyneaux

Eduardo P Oliveira

James Walter

Shao Liu

Oliver M Williams

Paul Ronald Barham

Dax H Hawkins

Mitchell Stephen Dernis

William F Schiefelbein

Bob Crocco

Sanjay Bhatt

James H Dooley, Iv

Jon Vincent

Molly Rose Suver

Scott Henson

Leo Shing

Morten Pedersen

Ryan O Rourke

Shawn Pile

Ryan O'Rourke

Steve Hodges

Otmar Hiliges

Hoang T Do

Lae-Hoon Kim

Xudong Ni

Iv James H Dooley

James T Walter

The disclosed system continuously refines a model used by an Automatic Speech Recognition (ASR) system to enable fast and accurate transcriptions of detected speech activity. The ASR system analyzes speech activity to generate text transcriptions and associated metrics (such as minimum Bayes risk and/or perplexity) that correspond to the quality of or confidence in each generated transcription. The system employs a filtering process to select certain text transcriptions based in part on one or more associated quality metrics. In addition, the system corrects for known systemic errors within the ASR system and provides a mechanism for manual review and correction of transcriptions. The system selects a subset of transcriptions based on factors including confidence score, and uses the selected subset of transcriptions to re-train the ASR model. By continuously retraining the ASR model, the system is able to provide ever faster and more accurate text transcriptions of detected speech activity.

Automatic speech recognition (ASR) model training

A system and method providing semi-private conversation using an area microphone between one local user in a group of local users and a remote user. The local and remote users may be in different physical environments, using devices coupled by a network. A conversational relationship is defined between a local user and a remote user. The local user's voice is isolated from other voices in the environment, and transmitted to the remote user. Directional output technology may be used to direct the local user's utterances to the remote user in the remote environment.

Semi-private communication in open environments

A multimedia entertainment system combines both gestures and voice commands to provide an enhanced control scheme. A user's body position or motion may be recognized as a gesture, and may be used to provide context to recognize user generated sounds, such as speech input. Likewise, speech input may be recognized as a voice command, and may be used to provide context to recognize a body position or motion as a gesture. Weights may be assigned to the inputs to facilitate processing. When a gesture is recognized, a limited set of voice commands associated with the recognized gesture are loaded for use. Further, additional sets of voice commands may be structured in a hierarchical manner such that speaking a voice command from one set of voice commands leads to the system loading a next set of voice commands.

Compound gesture-speech commands

A system and method for monitoring telephone calls to detect fraudulent activity and take corrective action is described. The system receives a group of telephone calls having associated call characteristics and analyzes the group of telephone calls to identify and store a first set of distributions of call characteristics that are indicative of normal activity, fraudulent activity, or indeterminate activity. The system receives one or more subsequent telephone calls to be analyzed. The system analyzes the received one or more telephone calls to identify a second set of distributions of call characteristics associated with the received telephone call. The system then compares the second set of distributions of call characteristics to the stored first set of distributions of call characteristics to assess a probability that the one or more received telephone calls represents normal, fraudulent, or indeterminate activity. If the assessed probability of fraudulent activity exceeds a threshold, the system takes appropriate corrective action, such a flagging the fraudulent call or withholding a financial incentive associated with the fraudulent call.

Identifying call characteristics to detect fraudulent call activity and take corrective action without using recording, transcription or caller ID

A facility and method for analyzing and classifying calls without transcription. The facility analyzes individual frames of an audio to identify speech and measure the amount of time spent in speech for each channel (e.g., caller channel, agent channel). Additional telephony metrics such as R-factor or MOS score and other metadata may be factored in as audio analysis inputs. The facility then analyzes the frames together as a whole and formulates a clustered-frame representation of a conversation to further identify dialog patterns and characterize call classification. Based on the data in the clustered-frame representation, the facility is able to make estimations of call classification. The correlation of dialog patterns to call classification may be utilized to develop targeted solutions for call classification issues, target certain advertising channels over others, evaluate advertising placements at scale, score callers, and to identify spammers.

System and method for analyzing and classifying calls without transcription

A system and method for monitoring telephone calls to detect call traffic pumping activity and take corrective action is described. The system receives a group of training telephone calls having associated call audio content and associated information, and the system analyzes the training telephone calls to generate and store a classification model that correlates call features and associations with a probability of call traffic pumping activity. The system receives a subsequent monitored telephone call to be analyzed. The system analyzes the monitored telephone call to identify features present in the audio content of the monitored telephone call and other associated information. The system then compares the features and associated information to the stored classification model in order to determine a probability that the monitored telephone call is associated with call traffic pumping activity. If the assessed probability of call traffic pumping activity exceeds a threshold, the system takes appropriate corrective action, such as terminating or flagging the monitored call.

Identifying call features and associations to detect call traffic pumping and take corrective action

A facility and method for analyzing and classifying calls without transcription via keyword spotting is disclosed. The facility uses a group of calls having known outcomes to generate one or more domain- or entity-specific grammars containing keywords and related information that are indicative of particular outcome. The facility monitors telephone calls by determining the domain or entity associated with the call, loading the appropriate grammar or grammars associated with the determined domain or entity, and tracking keywords contained in the loaded grammar or grammars that are spoken during the monitored call, along with additional information. The facility performs a statistical analysis on the tracked keywords and additional information to determine a classification for the monitored telephone call.

System and method for analyzing and classifying calls without transcription via keyword spotting

A system and method for monitoring telephone calls to detect fraudulent activity and take corrective action is described. The system receives a first group of telephone calls having associated voice characteristics and analyzes the first group of telephone calls to identify and store a first set of distributions of voice characteristics that are indicative of normal activity, fraudulent activity, or indeterminate activity. The system receives a second group of telephone calls to be analyzed. The system analyzes the second group of telephone calls to identify a second set of distributions of voice characteristics associated with the second group of telephone calls. The system then compares the second set of distributions of voice characteristics to the stored first set of distributions of voice characteristics to assess a probability that one or more telephone calls in the second group of telephone calls represents normal, fraudulent, or indeterminate activity. If the assessed probability of fraudulent activity exceeds a threshold, the system takes appropriate corrective action, such a flagging the fraudulent call or withholding a financial incentive associated with the fraudulent call.

Analyzing voice characteristics to detect fraudulent call activity and take corrective action without using recording, transcription or caller ID

An audio/visual system (e.g., such as an entertainment console or other computing device) plays a base audio track, such as a portion of a pre-recorded song or notes from one or more instruments. Using a depth camera or other sensor, the system automatically detects that a user (or a portion of the user) enters a first collision volume of a plurality of collision volumes. Each collision volume of the plurality of collision volumes is associated with a different audio stem. In one example, an audio stem is a sound from a subset of instruments playing a song, a portion of a vocal track for a song, or notes from one or more instruments. In response to automatically detecting that the user (or a portion of the user) entered the first collision volume, the appropriate audio stem associated with the first collision volume is added to the base audio track or removed from the base audio track.

Interactive content creation

A method for providing three-dimensional audio is provided. The method includes receiving a depth map imaging a scene from a depth camera and recognizing a human subject present in the scene. The human subject is modeled with a virtual skeleton comprising a plurality of joints defined with a three-dimensional position. A world space ear position of the human subject is determined based on the virtual skeleton. Furthermore, a target world space ear position of the human subject is determined. The target world space ear position is the world space position where a desired audio effect can be produced via an acoustic transducer array. The method further includes outputting a notification representing a spatial relationship between the world space ear position and the target world space ear position.

Three-dimensional audio sweet spot feedback

Growing community of inventors

Bellevue, WA, United States of America

Jason S Flaks