Adobe, Inc.

New York University

Justin Salamon

Aseem Agarwala

Lubomira A Dontcheva

Xue Bai

Hijung Shin

Joel Richard Brandt

Dingzeyu Li

Cristin Ailidh Fraser

Seth John Walker

Jovan Popović

Joy Oakyung Kim

Nicholas J Bryan

Zeyu Jin

Timothy Jeewun Ganter

Jui-hsien Wang

Bryan Russell

Jongpil Lee

Franck Dernoncourt

Jiyoung Lee

Fabian David Caba Heilbron

Carlos Augusto Dietrich

Karren Yang

Haoran Cai

Claudio T Silva

Yu Wang

Arvi Gjoka

Jorge Piazentin Ono

Didac Suris Coll-vinent

Oriol Nieto-Caballero

Some aspects of the technology described herein employ a neural network with an efficient and lightweight architecture to perform spoken language recognition. Given an audio signal comprising speech, features are generated from the audio signal, for instance, by converting the audio signal to a normalized spectrogram. The features are input to the neural network, which has one or more convolutional layers and an output activation layer. Each neuron of the output activation layer corresponds to a language from a set of language and generates an activation value. Based on the activations values, an indication of zero or more languages from the set of languages is provided for the audio signal.

Spoken language recognition

Embodiments are disclosed for correlating video sequences and audio sequences by a media recommendation system using a trained encoder network. In particular, in one or more embodiments, the disclosed systems and methods comprise receiving a training input including a media sequence, including a video sequence paired with an audio sequence, segmenting the media sequence into a set of video sequence segments and a set of audio sequence segments, extracting visual features for each video sequence segment and audio features for each audio sequence segment, generating, by transformer networks, contextualized visual features from the extracted visual features and contextualized audio features from the extracted audio features, the transformer networks including a visual transformer and an audio transformer, generating predicted video and audio sequence segment pairings based on the contextualized visual and audio features, and training the visual transformer and the audio transformer to generate the contextualized visual and audio features.

Self-supervised audio-visual learning for correlating music and video

Embodiments of the present invention provide systems, methods, and computer storage media for music-aware speaker diarization. In an example embodiment, one or more audio classifiers detect speech and music independently of each other, which facilitates detecting regions in an audio track that contain music but do not contain speech. These music-only regions are compared to the transcript, and any transcription and speakers that overlap in time with the music-only regions are removed from the transcript. In some embodiments, rather than having the transcript display the text from this detected music, a visual representation of the audio waveform is included in the corresponding regions of the transcript.

Music-aware speaker diarization for transcripts and text-based video editing

A method for detecting a cue (e.g., a visual cue or a visual cue combined with an audible cue) occurring together in an input video includes: presenting a user interface to record an example video of a user performing an act including the cue; determining a part of the example video where the cue occurs; applying a feature of the part to a neural network to generate a positive embedding; dividing the input video into a plurality of chunks and applying a feature of each chunk to the neural network to generate a plurality of negative embeddings; applying a feature of a given one of the chunks to the neural network to output a query embedding; and determining whether the cue occurs in the input video from the query embedding, the positive embedding, and the negative embeddings.

Automatic recognition of visual and audio-visual cues

Embodiments of the present invention provide systems, methods, and computer storage media for identifying candidate boundaries for video segments, video segment selection using those boundaries, and text-based video editing of video segments selected via transcript interactions. In an example implementation, boundaries of detected sentences and words are extracted from a transcript, the boundaries are retimed into an adjacent speech gap to a location where voice or audio activity is a minimum, and the resulting boundaries are stored as candidate boundaries for video segments. As such, a transcript interface presents the transcript, interprets input selecting transcript text as an instruction to select a video segment with corresponding boundaries selected from the candidate boundaries, and interprets commands that are traditionally thought of as text-based operations (e.g., cut, copy, paste) as an instruction to perform a corresponding video editing operation using the selected video segment.

Video segment selection and editing using transcript interactions

Embodiments are directed to a snap point segmentation that defines the locations of selection snap points for a selection of video segments. Candidate snap points are determined from boundaries of feature ranges of the video indicating when instances of detected features are present in the video. In some embodiments, candidate snap point separations are penalized for being separated by less than a minimum duration corresponding to a minimum pixel separation between consecutive snap points on a video timeline. The snap point segmentation is computed by solving a shortest path problem through a graph that models different snap point locations and separations. When a user clicks or taps on the video timeline and drags, a selection snaps to the snap points defined by the snap point segmentation. In some embodiments, the snap points are displayed during a drag operation and disappear when the drag operation is released.

Snap point video segmentation identifying selection snap points for a video

Embodiments are directed to interactive tiles that represent video segments of a segmentation of a video. In some embodiments, each interactive tile represents a different video segment from a particular video segmentation (e.g., a default video segmentation). Each interactive tile includes a thumbnail (e.g., the first frame of the video segment represented by the tile), some transcript from the beginning of the video segment, a visualization of detected faces in the video segment, and one or more faceted timelines that visualize a category of detected features (e.g., a visualization of detected visual scenes, audio classifications, visual artifacts). In some embodiments, interacting with a particular interactive tile navigates to a corresponding portion of the video, adds a corresponding video segment to a selection, and/or scrubs through tile thumbnails.

Interacting with semantic video segments through interactive tiles

Embodiments are directed to a thumbnail segmentation that defines the locations on a video timeline where thumbnails are displayed. Candidate thumbnail locations are determined from boundaries of feature ranges of the video indicating when instances of detected features are present in the video. In some embodiments, candidate thumbnail separations are penalized for being separated by less than a minimum duration corresponding to a minimum pixel separation (e.g., the width of a thumbnail) between consecutive thumbnail locations on a video timeline. The thumbnail segmentation is computed by solving a shortest path problem through a graph that models different thumbnail locations and separations. As such, a video timeline is displayed with thumbnails at locations on the timeline defined by the thumbnail segmentation, with each thumbnail depicting a portion of the video associated with the thumbnail location.

Thumbnail video segmentation identifying thumbnail locations for a video

The sport data tracking systems available today are based on specialized hardware to detect and track targets on the field. While effective, implementing and maintaining these systems pose a number of challenges, including high cost and need for close human monitoring. On the other hand, the sports analytics community has been exploring human computation and crowdsourcing in order to produce tracking data that is trustworthy, cheaper and more accessible. However, state-of-the-art methods require a large number of users to perform the annotation, or put too much burden into a single user. Example methods, systems and user interfaces that facilitate the creation of tracking data sequences of events (e.g., plays of baseball games) by warm-starting a manual annotation process using a vast collection of historical data are described.

Reducing human interactions in game annotation

In implementations of searching for music, a music search system can receive a music search request that includes a music file including music content. The music search system can also receive a selected musical attribute from a plurality of musical attributes. The music search system includes a music search application that can generate musical features of the music content, where a respective one or more of the musical features correspond to a respective one of the musical attributes. The music search application can then compare the musical features that correspond to the selected musical attribute to audio features of audio files, and determine similar audio files to the music file based on the comparison of the musical features to the audio features of the audio files.

Searching for music

Certain embodiments involve techniques for automatically identifying sounds in an audio recording that match a selected sound. An audio search and editing system receives the audio recording and preprocesses the audio recording into audio portions. The audio portions are provided as a query to the neural network that includes a trained embedding model used to analyze the audio portions in view of the selected sound to estimate feature vectors. The audio search and editing system compares the feature vectors for the audio portions against the feature vector for the selected sound and the feature vector for the negative samples to generate an audio score that is a numerical representation of the level of similarity between the audio portion and the selected sound and uses the audio scores to classify the audio portions into a first class of matching sounds and a second class of non-matching sounds.

Automated sound matching within an audio recording

A computer system is trained to understand audio-visual spatial correspondence using audio-visual clips having multi-channel audio. The computer system includes an audio subnetwork, video subnetwork, and pretext subnetwork. The audio subnetwork receives the two channels of audio from the audio-visual clips, and the video subnetwork receives the video frames from the audio-visual clips. In a subset of the audio-visual clips the audio-visual spatial relationship is misaligned, causing the audio-visual spatial cues for the audio and video to be incorrect. The audio subnetwork outputs an audio feature vector for each audio-visual clip, and the video subnetwork outputs a video feature vector for each audio-visual clip. The audio and video feature vectors for each audio-visual clip are merged and provided to the pretext subnetwork, which is configured to classify the merged vector as either having a misaligned audio-visual spatial relationship or not. The subnetworks are trained based on the loss calculated from the classification.

Growing community of inventors

San Francisco, CA, United States of America

Justin Salamon