Back to results
US 7,734,460 B2Granted
Time asynchronous decoding for long-span trajectory model
AssigneeMicrosoft Corporation
InventorsDong Yu, Li Deng, Alejandro Acero
Life of this patent
Estimating the term…
Abstract
A time-asynchronous lattice-constrained search algorithm is developed and used to process a linguistic model of speech that has a long-contextual-span capability. In the algorithm, nodes and links in the lattices developed from the model are expanded via look-ahead. Heuristics as utilized by a search algorithm are estimated. Additionally, pruning strategies can be applied to speed up the search.
Claims13
A computer-implemented method of processing speech using a computer with a processor, comprising:
- receiving a speech signal;
- generating, with the processor, a lattice having a plurality of possible sequences, each sequence having a plurality of nodes representing a plurality of hypothesized linguistic units and transitions connecting the linguistic units for the received speech signal;
- converting, with the processor, each of the plurality of nodes to an expanded node such that each expanded node is representative of previous nodes in the sequence and future nodes in the sequence;
- generating, with the processor, scores for each of the plurality of expanded nodes;
- storing the scores for the expanded nodes;
- identifying, with the processor, a second expanded node in the lattice that has similar previous nodes and similar future nodes to a given one of the expanded nodes;
- generating a score for the second expanded node using the stored score for the given one of the expanded nodes;
- generating scores for each possible sequence in the plurality of possible sequences based on the scores for the expanded nodes and the second expanded node;
- comparing, with the processor, the scores for each possible sequence;
- and
- determining, with the processor, a best possible sequence from the plurality of possible sequences based on the scores for each possible sequence.
8 more elements
A speech recognizer, comprising:
- an acoustic model generating a lattice representing a plurality of possible sequences of hypothesized linguistic units and transitions connecting the linguistic units for a speech signal, each lattice having a plurality of sequences of nodes representing the hypothesized linguistic units and a plurality of arcs between the nodes representing the transitions in the possible sequence of hypothesized linguistic units;
- a decoder coupled to the acoustic model, converting each of the plurality of nodes to an expanded node, each expanded node representing previous nodes in a sequence of nodes and future nodes in the sequence of nodes and generating scores for transitions between linguistic units in the plurality of possible sequences, generating scores for the linguistic units based on the scores of previous transitions in the sequence, generating scores for each possible sequence in the plurality of possible sequences based on the scores for the linguistic units in each possible sequence and comparing the scores for each possible sequence;
- and
- wherein generating scores for transitions, generating scores for the linguistic units, and generating scores for each possible sequence comprises:
- generating a score for each expanded node;
- storing scores for each expanded node;
- identifying a set of second expanded nodes in the lattice, each second expanded node having similar previous nodes and future nodes to an associated one of the expanded nodes;
- for each given second expanded node, using the stored score for the associated expanded node to score the given second expanded node;
- and
- a computer processor, activated by the decoder, facilitating generation of scores.
7 more elements
A computer-implemented method of processing a model using a computer with a processor, comprising:
- accessing, with the processor, a lattice having a plurality of possible sequences, each sequence having a plurality of nodes representing linguistic units and a plurality of arcs representing transitions between the linguistic units;
- and
- converting, with the processor, each of the plurality of nodes for each sequence in the lattice to an expanded node such that each expanded node includes an indication of previous nodes in the sequence and an indication of future nodes in the sequence;
- generating a score, with the processor, for a given expanded node;
- storing the score, with the processor, for the given expanded node;
- identifying, with the processor, a second expanded node in the lattice that has similar previous nodes and similar future nodes to the given expanded node;
- and
- using the stored score for the given expanded node to score, with the processor, the second expanded node.
5 more elements
Dependencies are inferred from the claim text.