Google Inc.

At&t Intellectual Property I, L.p.

At+t Corp.

Lucent Technologies Inc.

Nuance Communications, Inc.

Michael Dennis Riley

Mehryar Mohri

Mark Charles Beutnagel

Johan Schalkwyk

Cyril Georges Luc Allauzen

Ciprian Ioan Chelba

Tomoyuki Nanno

Corinna Cortes

Andrew M Rosenberg

Murat Saraclar

Vincent Goffin

Martin Jansche

Terry Tai

Fernando Carlos Pereira

Gaku Ueda

Edward Oscar Benson

Zhiyan Liu

Fernando Carlos Neves Pereira

Pedro Jose Moreno Mengibar

Petar Aleksic

Michael H Cohen

Brandon M Ballinger

Eugene Weinstein

Richard William Sproat

Leonid Velikovich

Nils Klarlund

Pi-Yu Chung

Boulos Harb

David Bjorn Roe

Anaya Misra

Pavel Golik

Alex Rudnick

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for tagging during speech recognition. A word lattice that indicates probabilities for sequences of words in an utterance is obtained. A conditional probability transducer that indicates a frequency that sequences of both the words and semantic tags for the words appear is obtained. The word lattice and the conditional probability transducer are composed to construct a word lattice that indicates probabilities for sequences of both the words in the utterance and the semantic tags for the words. The word lattice that indicates probabilities for sequences of both the words in the utterance and the semantic tags for the words is used to generate a transcription that includes the words in the utterance and the semantic tags for the words.

Semantic model for tagging of word lattices

A speech synthesis process can record concatenation costs of unit sequential pairs to a concatenation cost database for speech synthesis by synthesizing speech from a text, identifying an acoustic unit sequential pair in the speech, searching for a concatenation cost for the acoustic unit sequential pair in a database using a hash table for the database, and when the concatenation cost is not found in the database, assigning a default value as the concatenation cost for the acoustic unit sequential pair.

Concatenation cost in speech synthesis for acoustic unit sequential pair using hash table and default concatenation cost

A system is configured to receive a first string corresponding to an interpretation of a natural-language user voice entry; provide a representation of the first string as feedback to the natural-language user voice entry; receive, based on the feedback, a second string corresponding to a natural-language corrective user entry, where the natural-language corrective user entry may correspond to a correction to the natural-language user voice entry; parse the second string into one or more tokens; determine at least one corrective instruction from the one or more tokens of the second string; generate, from at least a portion of each of the first and second strings and based on the at least one corrective instruction, candidate corrected user entries; select a corrected user entry from the candidate corrected user entries; and output the selected, corrected user entry.

Natural language correction for speech input

A speech synthesis system can record concatenation costs of most common acoustic unit sequential pairs to a concatenation cost database for speech synthesis by synthesizing speech from a text, identifying a most common acoustic unit sequential pair in the speech, assigning a concatenation cost to the most common acoustic sequential pair, and recording the concatenation cost of the most common acoustic sequential pair to a concatenation cost database.

Recording concatenation costs of most common acoustic unit sequential pairs to a concatenation cost database for speech synthesis

A data processing apparatus is configured to receive a first string related to a natural-language voice user entry and a second string including at least one natural-language refinement to the user entry; parse the first string into a first set of one or more tokens and the second string into a second set of one or more tokens; determine at least one refining instruction from the second set of one or more tokens; generate, from at least a portion of each of the first string and the second string and based on the at least one refining instruction, a group of candidate refined user entries; select a refined user entry from the group of candidate refined user entries; and output the selected, refined user entry.

Natural language refinement of voice and text entry

A statistical model may be created that relates human ratings of documents to objective signals generated from the documents, search queries, and/or other information (e.g., query logs). The model can then be used to predict human ratings/rankings for new documents/search query pairs. These predicted ratings can be used to, for example, refine rankings from a search engine or assist in evaluating or monitoring the efficacy of a search engine system.

Prediction of human ratings or rankings of information retrieval quality

The subject matter of this specification can be implemented in a computer-implemented method that includes receiving utterances and transcripts thereof. The method includes analyzing the utterances and transcripts to determine certain attributes, such as distances between prosodic contours for pairs of utterances. A model can be generated that can be used to estimate a distance between a determined prosodic contour for a received utterance and an unknown prosodic contour for a synthesized utterance when given a distance between attributes for text associated with the received utterance and the synthesized utterance.

Generating prosodic contours for synthesized speech

Methods, computer program products and systems are described for speech-to-text conversion. A voice input is received from a user of an electronic device and contextual metadata is received that describes a context of the electronic device at a time when the voice input is received. Multiple base language models are identified, where each base language model corresponds to a distinct textual corpus of content. Using the contextual metadata, an interpolated language model is generated based on contributions from the base language models. The contributions are weighted according to a weighting for each of the base language models. The interpolated language model is used to convert the received voice input to a textual output. The voice input is received at a computer server system that is remote to the electronic device. The textual output is transmitted to the electronic device.

Context based language model selection

Systems and methods for identifying music in a noisy environment are described. One of the methods includes receiving audio segment data. The audio segment data is generated from the portion that is captured in the noisy environment. The method further includes generating feature vectors from the audio segment data, identifying phonemes from the feature vectors, and comparing the identified phonemes with pre-assigned phoneme sequences. Each pre-assigned phoneme sequence identifies a known music piece. The method further includes determining an identity of the music based on the comparison.

Systems and methods for identifying music in a noisy environment

A speech synthesis system can select recorded speech fragments, or acoustic units, from a very large database of acoustic units to produce artificial speech. When a pair of acoustic units in the database does not have an associated concatenation cost, the system assigns a default concatenation cost. The system then synthesizes speech, identifies the acoustic unit sequential pairs generated and their respective concatenation costs, and stores those concatenation costs likely to occur.

Speech synthesis from acoustic units with default values of concatenation cost

A system extracts a pair that includes a keyword candidate and information associated with a document from multiple documents, and calculates a frequency that the keyword candidate appears in search queries and a frequency that the pair appears in the multiple documents. The system also determines whether the keyword candidate is a keyword for a category based on the calculated frequencies, and associates the keyword with the document if the keyword candidate is the keyword for the category.

Keywords associated with document categories

Systems and methods for identifying the N-best strings of a weighted automaton. A potential for each state of an input automaton to a set of destination states of the input automaton is first determined. Then, the N-best paths are found in the result of an on-the-fly determinization of the input automaton. Only the portion of the input automaton needed to identify the N-best paths is determinized. As the input automaton is determinized, a potential for each new state of the partially determinized automaton is determined and is used in identifying the N-best paths of the determinized automaton, which correspond exactly to the N-best strings of the input automaton.

Systems and methods for determining the N-best strings

A system identifies a document that includes an address and locates business information in the document. The system assigns a confidence score to the business information, where the confidence score relates to a probability that the business information is associated with the address. The system determines whether to associate the business information with the address based on the assigned confidence score.

Local item extraction

The subject matter of this specification can be implemented in, among other things, a computer-implemented method including receiving text to be synthesized as a spoken utterance. The method includes analyzing the received text to determine attributes of the received text and selecting one or more utterances from a database based on a comparison between the attributes of the received text and attributes of text representing the stored utterances. The method includes determining, for each utterance, a distance between a contour of the utterance and a hypothetical contour of the spoken utterance, the determination based on a model that relates distances between pairs of contours of the utterances to relationships between attributes of text for the pairs. The method includes selecting a final utterance having a contour with a closest distance to the hypothetical contour and generating a contour for the received text based on the contour of the final utterance.

A speech synthesis system can select recorded speech fragments, or acoustic units, from a very large database of acoustic units to produce artificial speech. The selected acoustic units are chosen to minimize a combination of target and concatenation costs for a given sentence. However, as concatenation costs, which are measures of the mismatch between sequential pairs of acoustic units, are expensive to compute, processing can be greatly reduced by pre-computing and caching the concatenation costs. Unfortunately, the number of possible sequential pairs of acoustic units makes such caching prohibitive. However, statistical experiments reveal that while about 85% of the acoustic units are typically used in common speech, less than 1% of the possible sequential pairs of acoustic units occur in practice. A method for constructing an efficient concatenation cost database is provided by synthesizing a large body of speech, identifying the acoustic unit sequential pairs generated and their respective concatenation costs, and storing those concatenation costs likely to occur. By constructing a concatenation cost database in this faction, the processing power required at run-time is greatly reduced with negligible effect on speech quality.

Growing community of inventors

New York, NY, United States of America

Michael Dennis Riley