Histogram based pre-pruning scheme for active HMMS
View Patent ↗Embodiments of the present invention include an acoustic processing device, a method for acoustic signal processing, and a speech recognition system. The speech processing device can include a processing unit, a histogram pruning unit, and a pre-pruning unit. The processing unit is configured to calculate one or more Hidden Markov Model (HMM) pruning thresholds. The histogram pruning unit is configured to prune one or more HMM states to generate one or more active HMM states. The pruning is based on the one or more pruning thresholds. The pre-pruning unit is configured to prune the one or more active HMM states based on an adjustable pre-pruning threshold. Further, the adjustable pre-pruning threshold is based on the one or more pruning thresholds.
1. An acoustic processing, device comprising:
an interface for receiving acoustic features obtained from an input analog signal representing an incoming voice signal;
a processing unit configured to calculate one or more current Hidden Markov Model (HMM) pruning thresholds;
a histogram pruning unit configured to prune one or more current HMM states based on the one or more pruning thresholds to generate one or more active HMM states;
a pre-pruning unit configured to prune the one or more active HMM states based on an adjustable pre-pruning threshold to generate active HMM states output of the pre-pruning unit, wherein the adjustable pre-pruning threshold is based on one or more prior pruning thresholds, and wherein the active HMM states output of the pre-pruning unit each indicate a phoneme of the incoming voice signal and are each associated with a HMM state score; and
a senone scoring unit configured to provide a senone score to the pre-pruning unit,
wherein the pre-pruning unit modifies the adjustable pre-pruning threshold based on the senone score
wherein the pre-pruning follows the histogram pruning for a current frame the incoming voice signal and precedes the histogram pruning for a next frame, and
wherein the interface transfers the phonemes and their associated HMM state scores to a further speech recognition stage and the further speech recognition stage generates recognized speech corresponding to the incoming voice signal.
2. The acoustic processing device according to claim 1 , further comprising:
a histogram generator configured to provide an HMM state score to the pre-pruning unit, wherein the pre-pruning unit modifies the adjustable pre-pruning threshold based on the HMM state score.
3. The acoustic processing device according to claim 2 , wherein the HMM state score comprises a best HMM state score from a previous frame of data.
4. The acoustic processing device according to claim 1 , wherein the senone score comprises a best senone score from a previous frame of data.
5. The acoustic processing device according to claim 1 , wherein the senone score comprises a best senone score from a current frame of data.
6. A method for acoustic signal processing comprising:
receiving acoustic features obtained from an input analog signal representing an incoming voice signal;
pruning one or more Hidden Markov Model (HMM) states based on one or more current HMM pruning thresholds to generate one or more active HMM states;
pre-pruning the one or more active HMM states based on an adjustable pre-pruning threshold to generate active HMM states output of the pre-pruning unit, wherein the adjustable pre-pruning threshold is based on one or more prior pruning thresholds, and wherein the active HMM states output of the pre-pruning unit each indicate a phoneme of the incoming voice signal and are each associated with a HMM state score;
calculating the adjustable pre-pruning threshold based on an HMM state score from a previous frame of data and a senone score, wherein the pre-pruning follows the pruning for a current frame the incoming voice signal and precedes the pruning for a next frame;
transferring the phonemes and their associated HMM state scores to a further speech recognition stage; and
generating recognized speech corresponding to the incoming voice signal at the further speech recognition stage.
7. The method according to claim 6 , wherein the HMM state score comprises a best HMM state score from the previous frame of data.
8. The method according to claim 6 , wherein the senone score comprises a best score from the current frame of data.
9. The method according to claim 6 , wherein the calculating the adjustable pre-pruning threshold is based on one or more senone scores from a previous frame of data.
10. The method according to claim 9 , wherein the one or more senone scores comprise a best senone score from the previous frame of data.
11. The method according to claim 6 , wherein the calculating the adjustable pre-pruning threshold is based on one or more senone scores from a current frame of data.
12. A speech recognition system comprising:
a central processing unit (CPU) with system memory, wherein the CPU receives an input analog signal representing an incoming voice signal and generates acoustic features corresponding to the incoming voice signal; and
an acoustic processing device coupled to the CPU, wherein the acoustic processing device comprises:
a processing unit configured to calculate one or more current Hidden Markov Model (HMM) pruning thresholds;
a histogram pruning unit configured to prune one or more HMM states based on the one or more current HMM pruning thresholds to generate one or more active HMM states;
a pre-pruning unit configured to prune the one or more active HMM states based on an adjustable pre-pruning threshold to generate active HMM states output of the pre-pruning unit, wherein the adjustable pre-pruning threshold is based on one or more prior pruning thresholds, wherein the pre-pruning follows the histogram pruning for a current frame the incoming voice signal and precedes the histogram pruning for a next frame,
wherein the active HMM states output of the pre-pruning unit each indicate a phoneme of the incoming voice signal and are each associated with a HMM state score; and
a senone scoring unit configured to receive the acoustic features and provide one or more senone scores to the pre-pruning unit,
wherein the pre-pruning unit modifies the adjustable pre-pruning threshold based on the one or more senone scores, and
wherein the acoustic processing device transfers the phonemes and their associated HMM state scores to the CPU and the CPU generates recognized speech corresponding to the incoming voice signal.
13. A system according to claim 12 , wherein the acoustic processing device further comprises:
a histogram generator configured to provide an HMM state score to the pre-pruning unit, wherein the pre-pruning unit modifies the adjustable pre-pruning threshold based on the HMM state score.
14. A system according to claim 12 , wherein the one or more senone scores comprise a best senone score from a previous frame of data or a current frame of data.