Speech recognition providing multiple outputs

Claim 2

In a speech recognition system of the type in which, in response to an input utterance characterized by a sequence of input acoustic segments, a path is dynamically decoded through a recognition network of sound characterizing arcs, predetermined ones of which connect at nodes which define permissible transitions, by matching sequential input acoustic segments to corresponding points along each possible arc and by measuring closeness of match to obtain a cost metric for reaching each node by the best path for each input time; a computer implemented method of generating an output network structure which is composed of a selection of said arcs explaining segments of the input utterance and interconnections thereof, said method comprising: for each arc, calculating and storing along with said node cost metrics, an arc score which represents the best path cost of reaching the last point on that arc at the corresponding time; when the end of the recognition network is reached, initiating a traceback procedure which comprises: assigning a minimal node out score to the end of the recognition network; for each successive node in said traceback procedure, selecting all incoming arcs having arc scores which, when combined with the node out score of the respective terminating node, do not exceed the best path score for the end of the recognition network by more than a preselected margin; and calculating, for each node which initiated selected arcs, a node out score which represents the cost of the best path from that node to the end of the recognition network, whereby the successively selected arcs and the joining nodes define the output network structure.

Claim 3

The method as set forth in claim 1 wherein each arc comprises a succession of spectral states.

Claim 4

The method as set forth in claim 2 wherein an input utterance is converted to a succession of spectral frames and said frames are matched to said states with said cost metric being a function of the closeness of match.

Claim 5

The method as set forth in claim 3 wherein said margin is a fixed offset in the cost metric.

Claim 6

In a speech recognition system of the type in which, in response to an input utterance characterized by a sequence of input acoustic segments, a path is dynamically decoded through a recognition network of sound characterizing arcs each of which comprises a succession of preselected spectral states, predetermined ones of said arcs being connected at nodes which define permissible transitions, by matching the spectral characteristics of sequential input acoustic segments to corresponding spectral states along each possible arc and by measuring closeness of match to obtain a cost metric for reaching each node by the best patch for each input time; a computer implemented method of generating an output network structure which is composed of a selection of said arcs and interconnections thereof together with corresponding costs, said method comprising: for each arc, calculating and storing along with said node cost metrics, an arc score which represents the best path cost of reaching the last point on that arc at the corresponding time and an arc intime which represents, for the best path, the time of leaving the previous arc; when the end of the recognition network is reached, initiating a traceback procedure which comprises: assigning a minimal node out score to the end of the recognition network; for each successive node encountered in said traceback procedure, selecting all incoming arcs having arc scores which, when combined with the node out score of the respective terminating node, do not exceed the best path score for the end of the recognition network by more than a preselected margin; and calculating, for each node which initiates selected arcs as identified by the respective arc intime, a node out score which represents the cost of the best path from that node to the end of the recognition network, whereby the successively selected arcs and the joining nodes define the output structure.

Claim 7

In a speech recognition system of the type in which, in response to an input utterance characterized by a sequence of input acoustic segments, a path is dynamically decoded through a recognition network of sound characterizing arcs each of which comprises a succession of preselected spectral states, predetermined ones of said arcs being connected at nodes which define permissible transitions, by matching the spectral characteristics of sequential input acoustic segments to corresponding spectral states along each possible arc and by measuring closeness of match to obtain a cost metric for reaching each node by the best patch for each input time; a computer implemented method of generating an output network structure which is composed of a selection of said arcs and interconnections thereof together with corresponding costs, said method comprising: for each arc, calculating and storing along with said node cost metrics, an arc score which represents the best path cost of reaching the last point on that arc at the corresponding time and an arc intime which represents, for the best path, the time of leaving the previous arc; when the end of the recognition network is reached, initiating a traceback procedure which comprises: assigning a minimal node out score to the end of the recognition network; for each successive node encountered in said traceback procedure, selecting all incoming arcs having arc scores which, when combined with the node out score of the respective terminating node, do not exceed the best path score for the end of the recognition network by more than a preselected margin; and calculating, for each node which initiates selected arcs as identified by the respective arc intime, a node out score which represents the cost of the best path from that node to the end of the recognition network, whereby the successively selected arcs and the joining nodes define the output structure.