At+t Corp.

At&t Intellectual Property I, L.p.

Lucent Technologies Inc.

Nuance Communications, Inc.

Mehryar Mohri

Michael Dennis Riley

Mark Charles Beutnagel

Mark-Jan Nederhof

Cyril Allauzen

Fernando Carlos Neves Pereira

Pi-Yu Chung

Fernando Carlos Pereira

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 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 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 context-free grammar can be represented by a weighted finite-state transducer. This representation can be used to efficiently compile that grammar into a weighted finite-state automaton that accepts the strings allowed by the grammar with the corresponding weights. The rules of a context-free grammar are input. A finite-state automaton is generated from the input rules. Strongly connected components of the finite-state automaton are identified. An automaton is generated for each strongly connected component. A topology that defines a number of states, and that uses active ones of the non-terminal symbols of the context-free grammar as the labels between those states, is defined. The topology is expanded by replacing a transition, and its beginning and end states, with the automaton that includes, as a state, the symbol used as the label on that transition. The topology can be fully expanded or dynamically expanded as required to recognize a particular input string.

Systems and methods for generating weighted finite-state automata representing grammars

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 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.

Methods and apparatus for rapid acoustic unit selection from a large speech corpus

An improved-removal method is disclosed that computes for any input weighted automaton A with-transitions an equivalent weighted automaton B with no-transitions. The method comprises two main steps. The first step comprises computing for each state 'p' of the automaton A its-closure. The second step in the method comprises modifying the outgoing transitions of each state 'p' by removing those labeled with. The method next comprises adding to the set of transitions leaving the state &#x201c;p&#x201d; non--transitions leaving each state &#x201c;q&#x201d; in the set of states reachable from &#x201c;p&#x201d; via a path labeled withwith their weights pre--multiplied by the-distance from state &#x201c;p&#x201d; to state &#x201c;q&#x201d; in the automaton A. State &#x201c;p&#x201d; is a final state if some state &#x201c;q&#x201d; within the set of states reachable from &#x201c;p&#x201d; via a path labeled withis final and the final weight 

System and method of epsilon removal of weighted automata and transducers

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. 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. By constructing a concatenation cost database in this fashion, the processing power required at run-time is greatly reduced with negligible effect on speech quality.

Finite-state transducers and weighted finite-state automata may not be determinizable. The twins property can be used to characterize the determinizability of such devices. For a weighted finite-state automaton or transducer, that weighted finite-state automaton or transducer and its inverse are intersected or composed, respectively. The resulting device is checked to determine if it has the cycle-identity property. If not, the original weighted finite-state automaton or transducer is not determinizable. For a weighted or unweighted finite-state transducer, that device is checked to determine if it is functional. If not, that device is not determinizable. That device is then composed with its inverse. The composed device is checked to determine if every edge in the composed device having a cycle-accessible end state meets at least one of a number of conditions. If so, the original device has the twins property. If the original device has the twins property, then it is determinizable.

Systems and methods for determining the determinizability of finite-state automata and transducers

A speech synthesis system can select recorded speech fragments, or acoustic units, from a 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. Concatenation costs are expensive to compute. Processing is reduced by pre-computing and caching the concatenation costs. The number of possible sequential pairs of acoustic units makes such caching prohibitive. 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.

Context-free grammars generally comprise a large number of rules, where each rule defines how a string of symbols is generated from a different series of symbols. While techniques for creating finite-state automata from the rules of context-free grammars exist, these techniques require an input grammar to be strongly regular. Systems and methods that convert the rules of a context-free grammar into a strongly regular grammar include transforming each input rule into a set of output rules that approximate the input rule. The output rules are all right- or left-linear and are strongly regular. In various exemplary embodiments, the output rules are output in a specific format that specifies, for each rule, the left-hand non-terminal symbol, a single right-hand non-terminal symbol, and zero, one or more terminal symbols. If the input context-free grammar rule is weighted, the weight of that rule is distributed and assigned to the output rules.

Systems and methods for regularly approximating context-free grammars through transformation

An improved &#x3b5;-removal method is disclosed that computes for any input weighted automaton A with &#x3b5;-transitions an equivalent weighted automaton B with no &#x3b5;-transitions. The method comprises two main steps. The first step comprises computing for each state 'p' of the automaton A its &#x3b5;-closure. The second step in the method comprises modifying the outgoing transitions of each state 'p' by removing those labeled with &#x3b5;. The method next comprises adding to the set of transitions leaving the state &#x201c;p&#x201d; non-&#x3b5;-transitions leaving each state &#x201c;q&#x201d; in the set of states reachable from &#x201c;p&#x201d; via a path labeled with &#x3b5; with their weights pre-{circle around (x)}multiplied by the &#x3b5;-distance from state &#x201c;p&#x201d; to state &#x201c;q&#x201d; in the automaton A. State &#x201c;p&#x201d; is a final state if some state &#x201c;q&#x201d; within the set of states reachable from &#x201c;p&#x201d; via a path labeled with &#x3b5; is final and the final weight 

Growing community of inventors

New York, NY, United States of America

Mehryar Mohri