System and method for providing large vocabulary speech processing based on fixed-point arithmetic

1. A method comprising:

representing arc costs associated with at least one finite-state transducer (FST) in fixed-point by converting floating-point costs of the at least one FST to fixed-point by quantization to yield fixed-point FST arc costs;

representing parameters associated with a hidden Markov model (HMM) in fixed-point to yield fixed-point HMM parameters; and

processing speech data via a processor in a speech recognition decoder using fixed-point arithmetic for the fixed-point FST arc costs and the fixed-point HMM parameters,

wherein at least the fixed-point HMM parameters are based on a Mahalanobis distance calculation, wherein a format for the fixed-point FST arc costs and the fixed-point HMM parameters is type Q-2e, wherein Q-2e is a fixed-point format with 2e bits representing a fractional part and e is a normalized error, and wherein the Mahalanobis distance calculation comprises computing Mahalanobis integer products for each Gaussian mean component of the parameters associated with the HMM, performing a right arithmetic shift for each of the Mahalanobis integer products to yield shifted integer products, and summing the shifted integer products.

2. The method of claim 1 , wherein the method is applied for large vocabulary automatic speech recognition.

3. The method of claim 1 , wherein the HMM is a phonetic context-dependent acoustic HMM.

4. The method of claim 1 , wherein the at least one FST is one of a single CLG FST or distinct C, L and G FSTs, wherein C FST is a model for co-articulation, L FST is a model for a lexicon and G FST is a model for a grammar.

5. The method of claim 1 , wherein representing parameters associated with the hidden Markov model (HMM) in fixed-point further comprises:

converting HMM state observation probability density functions to fixed-point format by quantization; and

computing HMM state likelihoods as Q-2e fixed-point numbers.

6. The method of claim 1 , wherein representing parameters associated with the hidden Markov model (HMM) in fixed-point further comprises:

converting HMM state duration probabilities to fixed-point by quantization using a fixed-point format of type Q-2e.

7. The method of claim 1 , further comprising:

computing, at the decoder, sentence hypothesis probabilities with fixed-point arithmetic as type Q-2e numbers.

8. A computing device that processes data associated with speech processing, the computing device comprising:

a processor;

a first module configured to cause the processor to represent arc costs associated with at least one finite-state transducer (FST) in fixed-point by converting floating-point costs of the at least one FST to fixed-point by quantization to yield fixed-point FST arc costs;

a second module configured to cause the processor to represent parameters associated with a hidden Markov model (HMM) in fixed-point to yield fixed-point HMM parameters; and

a third module configured to cause the processor to process speech data in a speech recognition decoder using fixed-point arithmetic for the fixed-point FST arc costs and the fixed-point HMM parameters,

wherein at least the fixed-point HMM parameters are based on a Mahalanobis distance calculation, wherein a format for the fixed-point FST arc costs and the fixed-point HMM parameters is type Q-2e, wherein Q-2e is a fixed-point format with 2e bits representing a fractional part and e is a normalized error, and wherein the Mahalanobis distance calculation comprises computing Mahalanobis integer products for each Gaussian mean component of the parameters associated with the HMM, performing a right arithmetic shift for each of the Mahalanobis integer products to yield shifted integer products, and summing the shifted integer products.

9. The computing device of claim 8 , wherein the computing device performs large vocabulary automatic speech recognition.

10. The computing device of claim 8 , wherein the HMM is a phonetic context-dependent acoustic HMM.

11. The computing device of claim 8 , wherein the at least one FST is one of a single CLG FST of distinct C, L and G FSTs, wherein C FST is a model for co-articulation, L FST is a model for a lexicon and G FST is a model for a grammar.

12. The computing device of claim 8 , wherein the second module configured to cause the processor to represent parameters associated with the hidden Markov model (HMM) in fixed-point further:

converts HMM state observation probability density functions to fixed-point format by quantization; and

computes HMM state likelihoods as Q-2e fixed-point numbers.

13. The computing device of claim 8 , wherein the second module configured to cause the processor to represent parameters associated with the hidden Markov model (HMM) in fixed-point further:

converts HMM state duration probabilities to fixed-point by quantization using a fixed-point format of type Q-2e.

14. The computing device of claim 8 , further comprising:

a fourth module configured to cause the processor to compute, at the decoder, sentence hypothesis probabilities with fixed-point arithmetic as type Q-2e numbers.

15. A non-transitory computer-readable medium storing instructions for controlling a computing device to perform steps comprising:

representing arc costs associated with at least one finite-state transducer (FST) in fixed-point by converting floating-point costs of the at least one FST to fixed-point by quantization to yield fixed-point FST arc costs;

representing parameters associated with a hidden Markov model (HMM) in fixed-point to yield fixed-point HMM parameters; and

processing speech data via a speech recognition decoder using fixed-point arithmetic for the fixed-point FST arc costs and the fixed-point HMM parameters,

wherein at least the fixed-point HMM parameters are based on a Mahalanobis distance calculation, wherein a format for the fixed-point FST arc costs and the fixed-point HMM parameters is type Q-2e, wherein Q-2e is a fixed-point format with 2e bits representing a fractional part and e is a normalized error, and wherein the Mahalanobis distance calculation comprises computing Mahalanobis integer products for each Gaussian mean component of the parameters associated with the HMM, performing a right arithmetic shift for each of the Mahalanobis integer products to yield shifted integer products, and summing the shifted integer products.