Block synchronous decoding

• US 20040176956A1
• Filed: 03/04/2003
• Published: 09/09/2004
• Est. Priority Date: 03/04/2003
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Odell, Julian J., Rockenbeck, William H.
• Granted Patent: US 7,529,671 B2
• Time in Patent Office: Days
• US Class Current: 704/256
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]

Abstract

A pattern recognition system and method are provided. Aspects of the invention are particularly useful in combination with multi-state Hidden Markov Models. Pattern recognition is effected by processing Hidden Markov Model Blocks. This block-processing allows the processor to perform more operations upon data while such data is in cache memory. By so increasing cache locality, aspects of the invention provide significantly improved pattern recognition speed.

Claims

1. A continuous pattern recognition system comprising:

an input device adapted to provide a digital representation of an input;

memory operably coupled to the input device to store the digital representation and a plurality of multi-state models relative to the digital representation;

a processor coupled to the input device, and the memory, the processor including a cache memory, and adapted to convert the digital representation into a plurality of time-sequenced frames; and

wherein the processor is adapted to generate an output of recognized patterns based upon processing the time-sequenced frames and blocks of the multi-state models stored in the cache memory.

2. The system of claim 1, wherein the multi-state models are three-state Hidden Markov Models.

3. The system of claim 1, wherein the input device is a microphone.

4. The system of claim 1, wherein the output of recognized patterns includes words.

5. The system of claim 1, wherein the system is embodied within a computer.

6. The system of claim 1, wherein the system is embodied within a mobile device.

7. A method of recognizing patterns in an input formed of time-sequenced frames, the method comprising:

modeling patterns with a plurality of multi-state Hidden Markov Models;

processing Hidden Markov Model Blocks (HMMBs) to recognize the modeled patterns among the time-sequenced frames to generate a sequence of recognized modeled patterns.

8. The method of claim 7, wherein the time-sequenced frame corresponds to speech.

9. The method of claim 7, wherein the multi-state Hidden Markov Models are tri-state Hidden Markov Models.

10. The method of claim 9, wherein each HMMB is a three by three rhombus when depicted on a state vs. time chart.

11. The method of claim 9, wherein processing each block includes interacting solely with system cache memory.

12. The method of claim 7, wherein processing the HMMBs includes accessing a language model.

13. A pattern recognizing method comprising:

representing an input as a series of time-sequenced frames; and

processing HMMBs and the series to generate an output sequence of recognized patterns corresponding to the input.

14. The method of claim 13, wherein the input corresponds to speech.

15. The method of claim 13, wherein each HMMB includes three states.

16. A method of recognizing patterns in an input formed of time-sequenced frames, the method comprising:

a) modeling patterns with a plurality of multi-state Hidden Markov Models;

b) processing a first point of a first multi-state Hidden Markov Model;

c) processing a second point of the first multi-state Hidden Markov Model, the second point differing from the first point in both state and time; and

d) processing remaining point of the first Hidden Markov Model, and points of the other of the plurality of multi-state Hidden Markov Models to recognize the modeled patterns among the time-sequenced frames to generate a sequence of recognized modeled patterns.

17. The method of claim 16, wherein the first point and the second point comprise an HMMB.