초록 ▼

A method is presented including selecting an initial beam width. The method also includes determining whether a value per frame is changing. A beam width is dynamically adjusted. The method further decides a speech input with the dynamically adjusted beam width. Also, a device is presented includin

A method is presented including selecting an initial beam width. The method also includes determining whether a value per frame is changing. A beam width is dynamically adjusted. The method further decides a speech input with the dynamically adjusted beam width. Also, a device is presented including a processor (420). A speech recognition component (610) is connected to the processor (420). A memory (410) is connected to the processor (420). The speech recognition component (610) dynamically adjusts a beam width to decode a speech input.

대표청구항 ▼

What is claimed is: 1. A method comprising: selecting an initial beam width; determining for a frame of input voice data a corresponding probability of each hypothesis in a set of hypotheses of a speech recognition model; determining a highest probability from among the determined probabilities for

What is claimed is: 1. A method comprising: selecting an initial beam width; determining for a frame of input voice data a corresponding probability of each hypothesis in a set of hypotheses of a speech recognition model; determining a highest probability from among the determined probabilities for the frame of input voice data; pruning a hypothesis from the set of hypotheses of the speech recognition model to create a pruned set of hypotheses, the pruning based on the initial beam width and the determined highest probability; determining a change in a likelihood of a best hypotheses per frame resulting from the creating the pruned set of hypotheses; dynamically adjusting the beam width based on the determined change in the likelihood of the best hypotheses per frame; and decoding a speech input with the dynamically adjusted beam width. 2. The method of claim 1, said decoding including pruning a set of active paths with the dynamically adjusted beam width. 3. The method of claim 1, said decoding further including using a hidden Markov model (HMM). 4. The method of claim 3, wherein the HMM is a Viterbi scoring search. 5. The method of claim 1, said dynamically adjusting including determining a first boundary coefficient and a second boundary coefficient based on heuristics. 6. A method comprising: determining an initial beam width; determining a plurality of threshold values for a plurality of active hypotheses of a speech recognition model, each of the active hypotheses corresponding to a respective probability for a frame of input voice data, the plurality of threshold values defining a first ranger; determining a current number of the plurality of active hypotheses; if the current number of active hypotheses is within the first range, determining a highest probability from among the probabilities of each of the active hypotheses for the frame of input voice data, dynamically adjusting the beam width based on the current number of active hypothesis and the threshold values, and selectively propagating one or more of the active hypotheses based on the determined highest probability and the dynamically adjusted beam width; and decoding a speech input based on the selectively propagated one or more hypotheses. 7. The method of claim 6 said decoding further including using a hidden Markov model (HMM). 8. The method of claim 7, wherein the HMM is a Viterbi scoring search. 9. The method of claim 6 said dynamically adjusting including determining a first boundary coefficient and a second boundary coefficient based on heuristics. 10. The method of claim 6, said determining the plurality of threshold 11. The method of claim 6 wherein the plurality of threshold values further defines a second range, the method further comprising: if the current number of active hypotheses is within the second range, selectively propagating one or more of the active hypotheses on an N-best basis. values comprises determining statistics on a sample of the speech input. 12. An apparatus comprising a machine-readable storage medium containing instructions which, when executed by a machine, cause the machine to perform operations comprising: selecting an initial beam width; determining for a frame of input voice data a corresponding probability of each hypothesis in a set of hypotheses of a speech recognition model; determining a highest probability from among the determined probabilities for the frame of input voice data; pruning a hypothesis from the set of hypotheses of the speech recognition model to create a pruned set of hypotheses, the pruning based on the initial beam width and the determined highest probability; determining a change in a likelihood of a best hypotheses per frame resulting from the creating the pruned set of hypotheses; adjusting the beam width dynamically based on the determined change in the likelihood of the best hypotheses per frame; and decoding a speech input with the dynamically adjusted beam width. 13. The apparatus of claim 12, said decoding including pruning a set of active paths with the dynamically adjusted beam_width. 14. The apparatus of claim 13, said decoding further including using a hidden Markov model (HMM). 15. The apparatus of claim 14, wherein the HMM is a Viterbi scoring search. 16. The apparatus of claim 14, said adjusting determining a first boundary coefficient and a second boundary coefficient based on heuristics. 17. An apparatus comprising a machine-readable storage medium containing instructions which, when executed by a machine, cause the machine to perform operations comprising: determining an initial beam width; determining a plurality of threshold values for a plurality of active hypotheses of a speech recognition model, each of the active hypotheses corresponding to a respective probability for a frame of input voice data, the plurality of threshold values defining a first range; determining a current number of the plurality of active hypotheses; if the current number of active hypotheses is within the first range, determining a highest probability from among the probabilities of each of the active hypotheses for the frame of input voice data, adjusting the beam width dynamically based on the current number of active hypothesis and the threshold values, and selectively propagating one or more of the active hypotheses based on the determined highest probability and the dynamically adjusted beam width; and decoding a speech input based on the selectively propagated one or more hypotheses. 18. The apparatus of claim 17, said decoding further including using a Viterbi scoring search. 19. The apparatus of claim 17, wherein the plurality of threshold values further defines a second range, the operations further comprising: if the current number of active hypotheses is within the second range, selectively propagating one or more of the active hypotheses on an N-best basis.