State output probability calculating method and apparatus for mixture distribution HMM

• US 20060229871A1
• Filed: 04/06/2006
• Published: 10/12/2006
• Est. Priority Date: 04/11/2005
• Status: Active Grant

• Current Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Original Assignee: Canon Kabushiki Kaisha (Canon Inc.)
• Inventors: Yamamoto, Hiroki, Yamada, Masayuki
• Granted Patent: US 7,813,925 B2
• Time in Patent Office: Days
• US Class Current: 704/256.600
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

When adjacent times or the small change of an observation signal is determined, a distribution which maximizes the output probability of a mixture distribution does not change at a high possibility. By using this fact, when obtaining the output probability of the mixture distribution HMM, a distribution serving as a maximum output probability is stored. When adjacent times or the small change of the observation signal is determined, the output probability of the stored distribution serves as the output probability of the mixture distribution. This can reduce the output probability calculation of other distributions when calculating the output probability of the mixture distribution, thereby reducing the calculation amount required for output probabilities.

Claims

1. A state output probability calculating method for a mixture distribution HMM, executed by a pattern recognition apparatus which performs pattern recognition of an observation signal by using a mixture distribution HMM in which an event of a recognition target stored in a memory is modeled, comprising:

a first output probability calculating step of calculating weighted output probabilities of all distributions belonging to a state for a first observation signal, obtaining a maximum value of the weighted output probabilities and a maximum distribution which outputs the maximum value, and defining the obtained maximum value as an output probability of a mixture distribution;

a maximum distribution storage step of storing, in the memory, information which specifies a maximum distribution obtained in the first output probability calculating step; and

a second output probability calculating step of, when a predetermined condition is satisfied, calculating a weighted output probability of the maximum distribution stored in the memory in the maximum distribution storage step, instead of the first output probability calculating step, for a second observation signal observed after the first observation signal, and defining the calculated weighted output probability as the output probability of the mixture distribution.

2. The method according to 1, wherein the predetermined condition is defined as a condition that a predetermined period of time has not elapsed from latest execution of the first output probability calculating step.

3. The method according to 1, wherein the predetermined condition is a condition that a distance between a current observation signal and a observation signal which is used in latest execution of the first output probability calculating step is less than a predetermined value.

4. The method according to claim 1, wherein the predetermined condition is a condition that a predetermined period of time has not elapsed from latest execution of the first output probability calculating step, and a distance between a current observation signal and a target observation signal in the latest execution of the first output probability calculating step is less than a predetermined value.

10. A program which causes a computer to execute an output probability calculating method defined in claim 1.

5. A state output probability calculating method for a mixture distribution HMM, executed by a pattern recognition apparatus which performs pattern recognition of an observation signal by using a mixture distribution HMM in which an event of a recognition target stored in a memory is modeled, comprising:

a first output probability calculating step of calculating weighted output probabilities of all distributions belonging to a state for a first observation signal, obtaining a maximum value of the weighted output probabilities and a maximum distribution which outputs the maximum value, and defining the obtained maximum value as an output probability of a mixture distribution;

a maximum distribution storage step of storing, in the memory, information which specifies a maximum distribution obtained in the first output probability calculating step;

a second output probability calculating step of calculating a weighted output probability of the maximum distribution stored in the memory in the maximum distribution storage step for a second observation signal observed after the first observation signal, and defining the calculated weighted output probability as an output probability of a mixture distribution; and

a control step of recalculating an output probability in the first output probability calculating step when the output probability calculated in the second output probability calculating step is not more than a predetermined threshold value.

11. A program which causes a computer to execute an output probability calculating method defined in claim 5.

6. An output probability calculating apparatus for calculating an output probability of a mixture distribution HMM state, comprising:

first output probability calculating means for calculating weighted output probabilities of all distributions belonging to states to obtain a maximum value thereof and a maximum distribution serving as a distribution which outputs the maximum value, and defining the obtained maximum value as an output probability of a mixture distribution;

storage means for storing information which specifies the maximum distribution obtained by said first output probability calculating means;

second output probability calculating means for calculating a weighted output probability of the maximum distribution stored in said storage means and defining the calculated weighted output probability as an output probability of a mixture distribution; and

control means for switching between said first output probability calculating means and said second output probability calculating means in accordance with whether a predetermined condition is satisfied.

8. A speech recognition apparatus including an output probability calculating apparatus defined in claim 6.

7. An output probability calculating apparatus for calculating an output probability of a mixture distribution HMM state, comprising:

first output probability calculating means for calculating weighted output probabilities of all distributions belong to states for a first observation signal, obtaining a maximum value of the weighted output probabilities and a maximum distribution serving as a distribution which outputs the maximum value, and defining the obtained maximum value as an output probability of a mixture distribution;

storage means for storing information which specifies a maximum distribution obtained by said first output probability calculating means;

second output probability calculating means for calculating a weighted output probability of the maximum distribution stored in said storage means for a second observation signal observed after the first observation signal, and defining the calculated weighted output probability as an output probability of a mixture distribution; and

control means for causing said first output probability calculating means to recalculate an output probability when the output probability calculated by said second output probability calculating means is not more than a predetermined threshold value.

9. A speech recognition apparatus including an output probability calculating apparatus defined in claim 7.