Google Inc.

Joel Eric Shor

Michele Mae Covell

George Dan Toderici

Nicholas M Johnston

David Charles Minnen

Sung Jin Hwang

Damien Vincent

Rahul Sukthankar

Saurabh Singh

Sean O'Malley

Alanna Foster Slocum

Aren Jansen

Sergio Guadarrama Cotado

Marco Tagliasacchi

Ronnie Maor

Oran Lang

Omry Tuval

Ira Shavitt

Felix De Chaumont Quitry

Dotan Emanuel

A method for determining synthetic speech includes receiving audio data characterizing speech in audio data obtained by a user device. The method also includes generating, using a trained self-supervised model, a plurality of audio features vectors each representative of audio features of a portion of the audio data. The method also includes generating, using a shallow discriminator model, a score indicating a presence of synthetic speech in the audio data based on the corresponding audio features of each audio feature vector of the plurality of audio feature vectors. The method also includes determining whether the score satisfies a synthetic speech detection threshold. When the score satisfies the synthetic speech detection threshold, the method includes determining that the speech in the audio data obtained by the user device comprises synthetic speech.

Self-supervised speech representations for fake audio detection

Examples relate to on-device non-semantic representation fine-tuning for speech classification. A computing system may obtain audio data having a speech portion and train a neural network to learn a non-semantic speech representation based on the speech portion of the audio data. The computing system may evaluate performance of the non-semantic speech representation based on a set of benchmark tasks corresponding to a speech domain and perform a fine-tuning process on the non-semantic speech representation based on one or more downstream tasks. The computing system may further generate a model based on the non-semantic representation and provide the model to a mobile computing device. The model is configured to operate locally on the mobile computing device.

Methods and systems for implementing on-device non-semantic representation fine-tuning for speech classification

Computing systems that provide a modularized infrastructure for training Generative Adversarial Networks (GANs) are provided herein. For example, the modularized infrastructure can include a lightweight library designed to make it easy to train and evaluate GANs. A user can interact with and/or build upon the modularized infrastructure to easily train GANs. The modularized infrastructure can include a number of distinct sets of code that handle various stages of and operations within the GAN training process. The sets of code can be modular. That is, the sets of code can be designed to exist independently yet be easily and intuitively combinable. Thus, the user can employ some or all of the sets of code or can replace a certain set of code with a set of custom-code while still generating a workable combination.

Computing systems with modularized infrastructure for training generative adversarial networks

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for image compression and reconstruction. A request to generate an encoded representation of an input image is received. The encoded representation of the input image is then generated. The encoded representation includes a respective set of binary codes at each iteration. Generating the set of binary codes for the iteration from an initial set of binary includes: for any tiles that have already been masked off during any previous iteration, masking off the tile. For any tiles that have not yet been masked off during any of the previous iterations, a determination is made as to whether a reconstruction error of the tile when reconstructed from binary codes at the previous iterations satisfies an error threshold. When the reconstruction quality satisfies the error threshold, the tile is masked off.

Stop code tolerant image compression neural networks

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for image compression and reconstruction. An image encoder system receives a request to generate an encoded representation of an input image that has been partitioned into a plurality of tiles and generates the encoded representation of the input image. To generate the encoded representation, the system processes a context for each tile using a spatial context prediction neural network that has been trained to process context for an input tile and generate an output tile that is a prediction of the input tile. The system determines a residual image between the particular tile and the output tile generated by the spatial context prediction neural network by process the context for the particular tile and generates a set of binary codes for the particular tile by encoding the residual image using an encoder neural network.

Tiled image compression using neural networks

Methods, and systems, including computer programs encoded on computer storage media for compressing data items with variable compression rate. A system includes an encoder sub-network configured to receive a system input image and to generate an encoded representation of the system input image, the encoder sub-network including a first stack of neural network layers including one or more LSTM neural network layers and one or more non-LSTM neural network layers, the first stack configured to, at each of a plurality of time steps, receive an input image for the time step that is derived from the system input image and generate a corresponding first stack output, and a binarizing neural network layer configured to receive a first stack output as input and generate a corresponding binarized output.

Image compression with recurrent neural networks

Growing community of inventors

Mountain View, CA, United States of America

Joel Eric Shor