Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for determining the number of speakers in a video and a corresponding audio using visual context. In one aspect, a method includes detecting within the video multiple speakers, determining a bounding box for each detected speaker that includes the detected person and objects within a threshold distance of the detected person in an image frame, determining a unique descriptor for that person based in part on image information depicting the objects within the bounding box, determining a cardinality of unique speakers in the video, providing to the speaker diarization system the cardinality of unique speakers.

Abstract: A method includes receiving an input text utterance to be synthesized into expressive speech having an intended prosody and a target voice and generating, using a first text-to-speech (TTS) model, an intermediate synthesized speech representation for the input text utterance. The intermediate synthesized speech representation possesses the intended prosody. The method also includes providing the intermediate synthesized speech representation to a second TTS model that includes an encoder portion and a decoder portion. The encoder portion is configured to encode the intermediate synthesized speech representation into an utterance embedding that specifies the intended prosody. The decoder portion is configured to process the input text utterance and the utterance embedding to generate an output audio signal of expressive speech that has the intended prosody specified by the utterance embedding and speaker characteristics of the target voice.

Abstract: A method includes obtaining training data including a plurality of training audio signals and corresponding transcripts. Each training audio signal is spoken by a target speaker in a first accent/dialect. For each training audio signal of the training data, the method includes generating a training synthesized speech representation spoken by the target speaker in a second accent/dialect different than the first accent/dialect and training a text-to-speech (TTS) system based on the corresponding transcript and the training synthesized speech representation. The method also includes receiving an input text utterance to be synthesized into speech in the second accent/dialect. The method also includes obtaining conditioning inputs that include a speaker embedding and an accent/dialect identifier that identifies the second accent/dialect.

Abstract: A method includes receiving an input text utterance to be synthesized into expressive speech having an intended prosody and a target voice and generating, using a first text-to-speech (TTS) model, an intermediate synthesized speech representation tor the input text utterance. The intermediate synthesized speech representation possesses the intended prosody. The method also includes providing the intermediate synthesized speech representation to a second TTS model that includes an encoder portion and a decoder portion. The encoder portion is configured to encode the intermediate synthesized speech representation into an utterance embedding that specifies the intended prosody. The decoder portion is configured to process the input text utterance and the utterance embedding to generate an output audio signal of expressive speech that has the intended prosody specified by the utterance embedding and speaker characteristics of the target voice.

Abstract: A method includes receiving an input text utterance to be synthesized into expressive speech having an intended prosody and a target voice and generating, using a first text-to-speech (TTS) model, an intermediate synthesized speech representation tor the input text utterance The intermediate synthesized speech representation possesses the intended prosody. The method also includes providing the intermediate synthesized speech representation to a second TTS model that includes an encoder portion and a decoder portion. The encoder portion is configured to encode the intermediate synthesized speech representation into an utterance embedding that specifies the intended prosody The decoder portion is configured to process the input text utterance and the utterance embedding to generate an output audio signal of expressive speech that has the intended prosody specified by the utterance embedding and speaker characteristics of the target voice.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating answer terms for scoring answer passages. In one aspect, a method includes accessing resource data describing a set of resources, identifying question phrases in the resources, for each identified question phrase in a resource, selecting in the resource a section of text subsequent to the question phrase as an answer, the answer having a plurality of terms, grouping the question phrases into groups of question phrases, and for each group: generating, from the terms of the answers for each question phrase in the group, answer terms and for each answer term, an answer term weight, and storing the answer terms and answer term weights in association with one or more queries.

Abstract: In general, aspects of the subject matter described in this specification can be embodied in methods that include the actions of receiving a search query initial input from a user, receiving a plurality of query auto-completions based on the search query initial input, receiving social graph data, the social graph data being specific to the user, for each query auto-completion of the plurality of query auto-completions, determining a ranking score, the ranking score being determined at least partially based on the social graph data, and transmitting instructions to display the plurality of query auto-completions to the user in a rank order that is determined based on ranking scores.

Abstract: Methods, systems, and apparatus are described that include receiving a sequence of one or more characters representing a first partial query entered into a search field on a computing device. A first list of query completions for the first partial query is provided for display on the computing device. One or more additional characters are then entered into the search field, where the sequence of characters and the additional characters cumulatively represent a second partial query. A second list of query completions are then obtained for the second partial query, and one or more query completions appearing in the first and the second lists are identified. Demotion scores for the identified query completions are then calculated which are used to demote the identified query completions to a lesser position within the second list, thereby forming a refined second list for display on the computing device.

Abstract: Technology described herein identifies and demotes query completions which divert users from their original search intent. The identified query completions may represent informational requests that are amusing diversions for some users, but which are not relevant to the actual information of interest for many users. For example, the identified query completions may represent information requests that some users may find humorous or entertaining, but which are not relevant to actual information of interest for many other users. The identification of these query completions may allow alternative query completions to be provided that can more likely assist users in finding the information they actually seek.

Abstract: Technology described herein enhances a user's search experience by providing refined search results that are relevant to a displayed document. Contextual search results are obtained which identify a list of documents responsive to a formulated query that is based on the user's search query, as well as one or more supplemental terms that are based on content in the displayed document during user entry of the search query. The contextual search results are then “refined” by re-ranking the documents in the list, based on the similarity between the user's original search query and terms in these documents. This re-ranking enables contextual search results to be provided that are also highly relevant to the user's informational need.

Abstract: Methods, systems and apparatus are described herein that include receiving a partial query entered into a search field on a computing device, and data indicating a document presented on the computing device during entry of the partial query. Search results are then obtained which identify respective lists of search results documents for corresponding candidate query completions in a set obtained for the partial query. Similarity scores are then calculated for candidate query completions based at least in part on a similarity between content in the displayed document and content in search result documents in the corresponding lists. One or more candidate query completions in the set are then selected as refined query completions for the partial query based at least in part on the similarity scores and provided for presentation by the computing device.

Abstract: In general, aspects of the subject matter described in this specification can be embodied in methods that include the actions of receiving a search query initial input from a user, receiving a plurality of query auto-completions based on the search query initial input, receiving social graph data, the social graph data being specific to the user, for each query auto-completion of the plurality of query auto-completions, determining a ranking score, the ranking score being determined at least partially based on the social graph data, and transmitting instructions to display the plurality of query auto-completions to the user in a rank order that is determined based on ranking scores.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for triggering music answer boxes. In one aspect, a method includes receiving a query, obtaining a plurality of search results responsive to the query, the search results being results from a search of web resources on the Internet, and determining from the plurality of search results that the query is a music query. Music data for a song responsive to the query is obtained, where the music data comprises a Uniform Resource Locator (URL) of song content of the song on the Internet. A music answer box is generated for the query, where the music answer box comprises the music data and a link to the URL of the song content, and the music answer box is provided in addition to search results.

Abstract: Methods, systems and apparatus are described herein that include obtaining metadata within a document, where the metadata comprises a sequence of terms. Tags are assigned to terms in the sequence of terms based at least in part on grammatical relationships between the terms, thereby forming a corresponding sequence of tags. A determination is made that the sequence of terms is grammatically correct based at least in part on tags within the corresponding sequence of tags. In response to the determination, the sequence of terms is stored as a query suggestion.

Abstract: Methods, systems and apparatus are described herein that include receiving a partial query entered into a search field on a computing device, and data indicating a document presented on the computing device during entry of the partial query. The presented document has been edited by a user of the computing device. Search results are then obtained which identify respective lists of search results documents for corresponding candidate query completions in a set obtained for the partial query. Similarity scores are then calculated for candidate query completions based at least in part on a similarity between content in the displayed document and content in search result documents in the corresponding lists. One or more candidate query completions in the set are then selected as refined query completions for the partial query based at least in part on the similarity scores and provided for presentation by the computing device.

Abstract: Methods, systems, and apparatus are described that include processing queries submitted by a plurality of users. A plurality of fresh queries is identified, where a fresh query is a query that has been submitted during a current time interval. The identified fresh queries are then reformulated into corresponding canonical representations using canonicalization rules. A group of fresh queries in the plurality of fresh queries having matching canonical representations are then selected. A group popularity score is then calculated for the group based at least in part on a number of times that one or more of the fresh queries in the group have been submitted during the current time interval. In response to a determination that the group popularity score satisfies a threshold popularity score, data is then stored identifying the fresh queries in the group of fresh queries as being permitted for use in determining a query suggestion.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for triggering music answer boxes. In one aspect, a method includes receiving a query, obtaining a plurality of search results responsive to the query, the search results being results from a search of web resources on the Internet, and determining from the plurality of search results that the query is a music query. Music data for a song responsive to the query is obtained, where the music data comprises a Uniform Resource Locator (URL) of song content of the song on the Internet. A music answer box is generated for the query, where the music answer box comprises the music data and a link to the URL of the song content, and the music answer box is provided in addition to search results.

Abstract: Methods and apparatus related to identifying one or more representative features of a live information source and ranking query suggestions based on the representative features. In some implementations the representative feature may be identified directly from the live information source. In some implementations the representative feature may be identified based on association of content of the live information source with the representative feature in a database. In some implementations additional factors related to the live information source may be utilized to rank the query suggestions.

Abstract: Technology is described herein for creating new, inferred queries that are suitable for use as query suggestions. The technology includes mapping different past queries that represent different information requests, but that share a number of terms having the same or similar form. An inferred query is created by replacing unshared terms in one mapped query with the unshared terms in another mapped query, based on the similarity between the unshared terms in the two mapped queries. The inferred query thus represents a new information request that is derived from the information requests of the past queries used to form it. The inferred query can then be provided as a query suggestion, which increases the likelihood of providing query suggestions that will assist users in finding the information they seek. In doing so, meaningful query suggestions can be provided to users seeking information outside the coverage of the past queries.

Abstract: The technology described identifies infrequently submitted past queries for use as query suggestions that are likely assist users in finding the information they seek. The technology includes filtering of infrequent queries by comparing canonical representations of the infrequent queries to canonical representations of popular queries. Canonical representations of infrequent queries are matched to canonical representations of popular queries; any infrequent queries are rejected from use as suggested queries if their canonical representation does not match that of any popular query. Selected infrequent queries can be stored as authorized for use by a subsequent computerized process in determining a query suggestion.