Baidu USA LLC

Vectra Networks, Inc.

University of California

Lawrence Livermore National Security, LLC

Ryan Prenger

Bryan Catanzaro

Sanjeev Satheesh

Shubhabrata Sengupta

Awni Hannun

Nicolas Beauchesne

Adam Coates

Erich Elsen

Christopher Fougner

Xu Han

Gregory Diamos

Jun Zhan

Jesse Engel

Karl Matthew Lynn

Zhenyao Zhu

Chong Wang

Andrew Ng

Mike Chrzanowski

Dario Amodei

Carl Case

Jared Casper

Jingdong Chen

Dani Yogatama

Sercan Omer Arik

Oliver Kourosh Tavakoli

Oskar Ibatullin

Andrew Gibiansky

John J Nitao

Joel Hestness

Markus Kliegl

Kofi Boakye

Tracy D Lemmond

William G Hanley

Rewon Child

Thomas Naselaris

Kendrick Kay

Joseph Wendell Guensche

Nathan C Perry

Paul Brandon Kidwell

Ron E Glaser

Erich Eisen

Jack L Gallant

Sengupta Shubhabrata

Presented herein are embodiments of state-of-the-art speech recognition systems developed using end-to-end deep learning. In embodiments, the model architecture is significantly simpler than traditional speech systems, which rely on laboriously engineered processing pipelines; these traditional systems also tend to perform poorly when used in noisy environments. In contrast, embodiments of the system do not need hand-designed components to model background noise, reverberation, or speaker variation, but instead directly learn a function that is robust to such effects. Neither a phoneme dictionary, nor even the concept of a 'phoneme,' is needed. Embodiments include a well-optimized recurrent neural network (RNN) training system that can use multiple GPUs, as well as a set of novel data synthesis techniques that allows for a large amount of varied data for training to be efficiently obtained. Embodiments of the system can also handle challenging noisy environments better than widely used, state-of-the-art commercial speech systems.

Deep learning models for speech recognition

Described herein are systems and methods for creating and using Convolutional Recurrent Neural Networks (CRNNs) for small-footprint keyword spotting (KWS) systems. Inspired by the large-scale state-of-the-art speech recognition systems, in embodiments, the strengths of convolutional layers to utilize the structure in the data in time and frequency domains are combined with recurrent layers to utilize context for the entire processed frame. The effect of architecture parameters were examined to determine preferred model embodiments given the performance versus model size tradeoff. Various training strategies are provided to improve performance. In embodiments, using only &#x2dc;230 k parameters and yielding acceptably low latency, a CRNN model embodiment demonstrated high accuracy and robust performance in a wide range of environments.

Convolutional recurrent neural networks for small-footprint keyword spotting

Presented herein are embodiments of state-of-the-art speech recognition systems developed using end-to-end deep learning. In embodiments, the model architecture is significantly simpler than traditional speech systems, which rely on laboriously engineered processing pipelines; these traditional systems also tend to perform poorly when used in noisy environments. In contrast, embodiments of the system do not need hand-designed components to model background noise, reverberation, or speaker variation, but instead directly learn a function that is robust to such effects. A phoneme dictionary, nor even the concept of a 'phoneme,' is needed. Embodiments include a well-optimized recurrent neural network (RNN) training system that can use multiple GPUs, as well as a set of novel data synthesis techniques that allows for a large amount of varied data for training to be efficiently obtained. Embodiments of the system can also handle challenging noisy environments better than widely used, state-of-the-art commercial speech systems.

Systems and methods for speech transcription

Embodiments of end-to-end deep learning systems and methods are disclosed to recognize speech of vastly different languages, such as English or Mandarin Chinese. In embodiments, the entire pipelines of hand-engineered components are replaced with neural networks, and the end-to-end learning allows handling a diverse variety of speech including noisy environments, accents, and different languages. Using a trained embodiment and an embodiment of a batch dispatch technique with GPUs in a data center, an end-to-end deep learning system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

End-to-end speech recognition

Deployed end-to-end speech recognition

A system and method for detecting malicious relay communications is disclosed. Network communications can be received and analyzed using such network components as a network switch. The received traffic can be parsed into sessions. Relay metadata can be extracted from the sessions and further be used to categorize the sessions into one or more types of relay metadata behaviors. Once a significant amount of sessions are detected an alarm may be triggered and/or alarm data may be generated for analysis by network security administrators.

Malicious relay detection on networks

Approaches for detecting network intrusions, such as malware infection, Trojans, worms, or bot net mining activities includes: identifying one or more threat detections in session datasets, the session datasets corresponding to network traffic from a plurality of hosts; determining a layered detection score, the layered detection score corresponding to a certainty score and threat score; determining a layered host score, the layered host score corresponding to a certainty score and threat score; and generating alarm data comprising the layered detection score and the layered host score. In some embodiments, the network traffic may be received passively through a network switch; for example, by 'tapping' the switch. Other additional objects, features, and advantages of the invention are described in the detailed description, figures and claims.

System and method for detecting network intrusions using layered host scoring

Decoding and reconstructing a subjective perceptual or cognitive experience is described. A first set of brain activity data produced in response to a first brain activity stimulus is acquired from a subject using a brain imaging device. An encoding model is used to convert the brain activity data into a corresponding set of predicted response values. A second set of brain activity data produced in response to a second brain activity stimulus is acquired from a subject and decoded using a decoding distribution derived from the encoding model, and the probability the second set of brain activity data corresponds to said predicted response values is determined. The second set of brain activity stimuli is then reconstructed based on the probability of correspondence between the second set of brain activity data and the predicted response values.

Apparatus and method for decoding sensory and cognitive information from brain activity

A detection engine may be implemented by receiving network traffic and processing the traffic into one or more session datasets. Sessions not initiated by an internal host may be discarded. The frequency between the communication packets from the internal host to external host may be grouped or processed into rapid-exchange instances. The number of rapid-exchange instances, the time intervals between them, and/or the rhythm and directions of the initiation of the instances may be analyzed to determine that a human actor is manually controlling the external host. In some embodiments, when it is determined that only one human actor is involved, alarm data may be generated that indicates that a network intrusion involving manual remote control has occurred or is underway.

Method and system for detecting external control of compromised hosts

An information extraction system and methods of operating the system are provided. In particular, an information extraction system for performing meta-extraction of named entities of people, organizations, and locations as well as relationships and events from text documents are described herein.

Information extraction system

Growing community of inventors

Oakland, CA, United States of America

Ryan Prenger