International Business Machines Corporation

Microsoft Technology Licensing, LLC

Baidu Online Network Technology (beijing) Co., Ltd.

Beijing Institute of Technology

Beijing Youzhuju Network Technology Co., Ltd.

Baidu Online Network Technology Co., Ltd.

Lei He

Raphael P Chancey

Eduardo T Kahan

Jian Luan

Max Leung

Jinyu Li

Yifan Gong

Zhe Yan

Sheng Zhao

Xue Wang

Jian Wu

Yulin Li

Guoli Ye

Edward C Lin

Kevin Lo

Yao Tian

Shifeng Pan

Daiqian Hu

Chunling Ma

Hongquan Yin

Wenning Wei

Eva Sharma

William Maxwell Gale

Yue Wang

Jian Ding

Yang Liu

Xin Peng Liu

Ke Zhang

Rui Zhao

Peng Ji

Yan Huang

Shaw-Ben Shi

Yong Zhao

Guo Qiang Hu

Xi Ning Wang

Kyle Gene Brown

Chunguang Xu

Yiu-Ming Leung

Meng Hong Tsai

Eric Nels Herness

Shun Xiang Yang

Cheng Luo

Andre Tost

Olga Ivanova

Art Sadovsky

Bo Yan

Pei Zhao

Venkat Narayanan

Smita Ojha

Tiancong Zhou

Chen Tian

Dingguo Xiao

Lei Niu

Yi Wang

Ashish Premaraj

Bo Wen Wei

Yujing Wang

Sixun Ma

Renyi Chen

Zhichao Yan

Haochen Xu

Wilson Man-Hong Li

Haoze Wu

Swati Singh

Ganyan Huang

Chang He Weng

Zhe Geng

Li Niu

Qiutao Wang

Hua Fu

Dong Bing Li

Wei Qing Wu

Systems, methods, and devices are provided for generating and using text-to-speech (TTS) data for improved speech recognition models. A main model is trained with keyword independent baseline training data. In some instances, acoustic and language model sub-components of the main model are modified with new TTS training data. In some instances, the new TTS training is obtained from a multi-speaker neural TTS system for a keyword that is underrepresented in the baseline training data. In some instances, the new TTS training data is used for pronunciation learning and normalization of keyword dependent confidence scores in keyword spotting (KWS) applications. In some instances, the new TTS training data is used for rapid speaker adaptation in speech recognition models.

Generating and using text-to-speech data for speech recognition models

The present disclosure relates to a method, apparatus, electronic device and storage medium for presentation of a matching scheme. The method comprises: acquiring a set of matching schemes; transmitting, to a first terminal, attribute information of matching schemes in the set of matching schemes, so that the first terminal displays attribute information of matching schemes comprised in the set of matching schemes; in response to a selection request for one of matching schemes transmitted by the first terminal, determining the selected matching scheme as a matching scheme to be explained; in response to a presentation request for the matching scheme to be explained that is transmitted by a second terminal, transmitting attribute information of the matching scheme to be explained to the second terminal, so that the second terminal presents a target information streaming page comprising at least a part of attribute information of the matching scheme to be explained.

Method, apparatus, electronic device, and storage medium for presentation of a matching scheme

A set of high level test logic is configured to include a set of insertion points. The high-level test logic can be controlled to insert test conditions into a data center configuration. It can also be configured to execute remediation actions that are to be taken, and validation actions to be performed in order to determine whether the remediation action works against the test conditions. Different instances of the high-level test logic can be configured for different environments and different test conditions.

Testing instrumentation for intrusion remediation actions

An ultrasonic detection and tensile calibration test method for bonding strength grade comprising bonding an upper substrate block to bonding groove(s) to form a theoretical bonding area, and applying a downward actual tensile force to a lower substrate block; obtaining an actual bonding area of the theoretical bonding area; calculating a first actual bonding strength by using the actual tensile force and the actual bonding area, and comparing the first actual bonding strength with a second actual bonding strength calculated to verify the correctness of the theoretical bonding area as a calibrated bonding strength; forming a bond strength table in which the theoretical bonding areas, the actual bonding areas and the first actual bonding strengths are in one-to-one correspondence; and using the actual bonding area to find the actual bonding strength corresponding to the actual bonding area from the bonding area bonding strength table.

Ultrasonic detection and tensile calibration test method for bonding strength grade

Systems are configured for generating spectrogram data characterized by a voice timbre of a target speaker and a prosody style of source speaker by converting a waveform of source speaker data to phonetic posterior gram (PPG) data, extracting additional prosody features from the source speaker data, and generating a spectrogram based on the PPG data and the extracted prosody features. The systems are configured to utilize/train a machine learning model for generating spectrogram data and for training a neural text-to-speech model with the generated spectrogram data.

System and method for cross-speaker style transfer in text-to-speech and training data generation

In one implementation, a first access behavior is determined. The first access behavior is for accessing a first dataset associated with a first tenant of a storage system, and the first dataset is stored in a first data node in the storage system. A second access behavior is determined. The second access behavior is for accessing a data entry that is to be added into a second dataset associated with a second tenant of the storage system, and the second dataset is stored in at least one data node in the storage system. In response to a deviation between the first access behavior and second access behavior being above a predefined threshold, the data entry is stored into the first data node.

Data management for multi-tenancy

Scheduling automated visual inspection tasks includes capturing an image of a component to be inspected. A visual inspection model is formed with a model engine as a composite model of utility modules and functional modules to perform visual inspection of the image of the component. An abstract processing workflow of the visual inspection model is derived with a scheduler including dependencies between the utility modules and the functional modules. Performance of each of the functional modules is profiled with the scheduler by testing performance with available hardware resources to produce a performance profile. Parallel instances of each of the functional modules in a branch of the abstract processing workflow are scheduled with the scheduler according to the dependencies and the performance profiles. An indication of defects in the component is produced by processing the visual inspection model according to the scheduled functional modules.

Flexible visual inspection model composition and model instance scheduling

Various embodiments of the present technology generally relate to systems and methods for self-healing services and automatic recovery of distribute systems. Some embodiments of the present technology leverage all the available synthetic, customer, client, server, support signals from various sources to intelligently and in real-time detect outages, root cause outages to recoverable targets (e.g., for auto recovery actions), identify the right engineering teams (e.g., for faster manual mitigation), and perform the appropriate recovery action (such as recycle service, reboot server, switch out a faulty rack) or other mitigation actions such as routing, collecting debug information, alerting to the right team, or alert suppression. Some embodiments separate signal monitoring and workflow coordination.

Automatic recovery engine with continuous recovery state machine and remote workflows

A computer-implemented method includes receiving a data stream and dividing the data stream into one or more data artefacts for one or more time intervals. The computer-implemented method further includes generating one or more contexts. Each of the one or more contexts are for at least one of the one or more time intervals. The computer-implemented method further includes storing each of the one or more data artefacts in shared storage. The shared storage is accessible for each of the one or more contexts. The computer-implemented method further includes, for each context of the one or more contexts: determining a dependency for the context; monitoring the context for a completion of said dependency; and responsive to the completion, releasing the dependency. The computer-implemented further includes removing those of the one or more data artefacts that do not belong to the dependency for any of the one or more contexts.

Efficiency for real-time data processing

A device controlling method, a client, a server and an intermediate device are provided. The device controlling method includes acquiring information on an intermediate device and information on a controlled device associated with the intermediate device; sending a resource downloading command carrying the information on the intermediate device which is to obtain a resource to a server so as to enable the intermediate device to obtain the resource via the server; and sending an operation instruction to the controlled device associated with the intermediate device so as to enable the controlled device to obtain the resource from the intermediate device to perform an operation according to the operation instruction.

Device controlling method, client, server and intermediate device

Multi-voice font interpolation is provided. A multi-voice font interpolation engine allows the production of computer generated speech with a wide variety of speaker characteristics and/or prosody by interpolating speaker characteristics and prosody from existing fonts. Using prediction models from multiple voice fonts, the multi-voice font interpolation engine predicts values for the parameters that influence speaker characteristics and/or prosody for the phoneme sequence obtained from the text to spoken. For each parameter, additional parameter values are generated by a weighted interpolation from the predicted values. Modifying an existing voice font with the interpolated parameters changes the style and/or emotion of the speech while retaining the base sound qualities of the original voice. The multi-voice font interpolation engine allows the speaker characteristics and/or prosody to be transplanted from one voice font to another or entirely new speaker characteristics and/or prosody to be generated for an existing voice font.

Voice font speaker and prosody interpolation

Outage detection in a cloud based service is provided using usage data based error signals. Usage data is collected from component of the cloud based service or client devices of the cloud based service based on customer actions on the cloud based service. The usage data is aggregated and normalized to generate an error signal from errors generated from a component of the cloud based service. An outage is detected from the error signal. An alert that includes information associated with the outage and one or more customers impacted by the outage is generated.

Detection of outage in cloud based service using usage data based error signals

Growing community of inventors

Beijing, China

Lei He