Utilizing multiple processing units for rapid training of hidden markov models

1. A method of optimizing the calculation of matching scores between phone states and acoustic frames across a matrix of a known, ordered progression of phone states of a transcript of at least one word aligned with an observed progression of acoustic frames within an utterance of the at least one word, the matrix having a plurality of cells, each cell associated with a characteristic acoustic frame and a characteristic phone state, the method comprising:

determining a first set of cells within a first row, the first row associated with a first phone state of the known, ordered progression of phone states of the transcript of the at least one word, wherein each cell of the first set of cells meets a threshold probability of matching the first phone state;

determining a second set of cells within a second row, the second row associated with a second phone state of the known, ordered progression of phone states of the transcript of the at least one word, wherein each cell of the second set of cells meets a threshold probability of matching the second phone state;

storing, on a local cache of a first core, the first phone state;

storing, on a local cache of a second core, the second phone state;

providing, to the first core, the first set of cells;

providing, to the second core, the second set of cells;

calculating, on the first core, matching scores of each characteristic state and characteristic observation of each cell of the first set of cells; and

calculating, on the second core, matching scores of each characteristic state and characteristic observation of each cell of the second set of cells, wherein the matching scores of the cells of the first set of cells and the matching scores of the cells of the second set of cells are calculated simultaneously.

2. The method of claim 1 , wherein each acoustic frame is a portion of the utterance.

3. The method of claim 1 , wherein a duration of each acoustic frame is the same for each cell.

4. The method of claim 1 , wherein each phone state is associated with a position within the progression of phone states.

5. The method of claim 4 , wherein the first phone state has a position that differs from the position of the second phone state.

6. The method of claim 5 , wherein the content of the first phone state matches the content of the second phone state.

7. The method of claim 1 , wherein the transcript of the at least one word comprises a plurality of words.

8. A method of optimizing the calculation of matching scores between phone states and acoustic frames across a matrix of a known, ordered progression of phone states of a transcript of at least one word aligned with an observed progression of acoustic frames within an utterance of the at least one word, the matrix having a plurality of cells, each cell associated with a characteristic acoustic frame and a characteristic phone state, the method comprising:

determining a first set of cells within a first row, the first row associated with a first phone state of the known, ordered progression of phone states of the transcript of the at least one word, wherein each cell of the first set of cells meets a threshold probability of matching the first phone state;

determining a second set of cells within a second row, the second row associated with a second phone state of the known, ordered progression of phone states of the transcript of the at least one word, wherein each cell of the second set of cells meets a threshold probability of matching the second phone state;

distributing storage of each phone state of the progression of phone states across a plurality of local caches on a chip;

providing, to a first core, the first set of cells associated with a first phone state that matches the phone state that is stored on the local cache that is associated with the first core;

providing, to a second core, the second set of cells associated with a second phone state that matches the phone state that is stored on the local cache that is associated with the second core;

calculating, on the first core, matching scores of each characteristic state and characteristic observation of each cell of the first set of cells; and

calculating, on the second core, matching scores of each characteristic state and characteristic observation of each cell of the second set of cells, wherein the matching scores of the cells of the first set of cells and the matching scores of the cells of the second set of cells are calculated simultaneously.

9. The method of claim 8 , wherein each phone state is distributed to a local cache based on an amount of storage space available on the local cache.

10. The method of claim 9 , wherein each phone state is distributed to the local cache having the greatest amount of available space at the time the phone state is distributed.

11. The method of claim 8 , wherein a plurality of phone states are stored on a local cache.

12. The method of claim 8 , wherein the transcript of the at least one word comprises a plurality of words.

13. A method of optimizing the calculation of matching scores between phone states and acoustic frames across a matrix of a known, ordered progression of phone states of a transcript of at least one word aligned with an observed progression of acoustic frames within an utterance of the at least one word, the matrix having a plurality of cells, each cell associated with a characteristic acoustic frame and a characteristic phone state, the method comprising:

determining a first set of cells within a first row, the first row associated with a first phone state of the known, ordered progression of phone states of the transcript of the at least one word, wherein each cell of the first set of cells meets a threshold probability of matching the first phone state;

storing, on a first core, a first phone state of the progression of phone states;

scheduling the calculation of matching scores of each cell of a first set of cells on the first core, where each cell of the first set of cells is associated with the first phone state;

calculating, on the first core, matching scores of each characteristic state and characteristic observation of each cell of the first set of cells, wherein the matching scores of the cells of the first set of cells are calculated simultaneously.

14. The method of claim 13 , wherein the first phone state is stored on a local cache of the first core.

15. The method of claim 13 , further comprising:

determining a second set of cells within a second row, the second row associated with a second phone state of the progression of phone states, wherein each cell of the second set of cells has a high probability of matching the second phone state;

scheduling the calculation of matching scores of each cell of a second row of cells on the first core, where each cell of the second row of cells is associated with the first phone state;

calculating, on the first core, matching scores of each characteristic state and characteristic observation of each cell of the second row of cells.

16. The method of claim 13 , wherein the plurality of cells comprise at least one anchor point.

17. The method of claim 16 , wherein a first anchor point comprises a cell associated with a first phone state of the progression of phone states, and the cell is further associated with a first acoustic frame of the progression of acoustic frames.

18. The method of claim 17 , wherein a second anchor point comprises a cell associated with a last phone state of the progression of phone states, and the cell is further associated with a last acoustic frame of the progression of acoustic frames.

19. The method of claim 13 , wherein the transcript of the at least one word comprises a plurality of words.