Pattern recognition apparatus

• US 5,129,002 A
• Filed: 11/05/1991
• Issued: 07/07/1992
• Est. Priority Date: 12/16/1987
• Status: Expired due to Term

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Tsuboka, Eiichi
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US07/787,747
• Time in Patent Office: 245 Days
• Field of Search: 381/42, 381/43, 395/2
• US Class Current: 704/246
• CPC Class Codes: G10L 15/144 Training of HMMs

Abstract

A pattern recognition apparatus using the hidden Markov model technique in which parameters for defining a mean vector representing the probability density function in each one of plural states for composing the hidden Markov model vary with time. Accordingly, the recognition precision may be enhanced when this apparatus is used in, for example, voice recognition.

Claims

1. A pattern recognition signal analysis apparatus for generating and storing an improved Hidden Markox Model (HMM), comprising:

feature extraction means for converting an input signal into a sequence of feature vectors, one for each state;

HMM creating means for making an HMM from said sequence of extracted feature vectors by calculating parameters for each said feature vector on the basis of previously-defined parameters to provide a mean vector representing the probability density function of each feature vector for each state, each said means vector varying with time within each state; and

means employing said HMM for recognizing an unknown pattern.

2. The pattern recognition apparatus of claim 1, wherein said parameters for defining each said mean vector for a state (i) comprise:

a neutral point vector μ

_i, a direction vector u_i, transition probability γ

_i, and a covariance matrix Σ

_i.

3. The pattern recognition apparatus of claim 2 further comprising:

means employing the created HMM for recognizing an unknown input signal.

4. The pattern recognition apparatus of claim 2 wherein said HMM creating means calculates said parameters from a training pattern set made of a plurality oflike training patterns.

5. The pattern recognition apparatus of claim 4 wherein said measn for recognizing an unknown input signal comprises:

means for converting said unknown input signal into an observed sequence of feature vectors, one for each state; and

means for calculating an observed mean vector representing the probability density function for each said feature vector for each state, each said observed mean vector varying with time within each state.

6. The pattern recognition apparatus of claim 5 wherein said means for recognizing further comprises:

means for calculating the likelihood in terms of a probability density function that said observed mean vector was generated by said HMM creating means.

7. A voice recognition apparatus using a plurality of Hidden Markov Models (HMM), each having a plurality of states, comprising:

feature extraction means for converting an input signal into a sequence of feature vectors;

HMM creating means for making a plurality of HMMs from the sequence of feature vectors, said HMM creating means utilizing at least one parameter for defining a mean vector in termsofa probability density function of the feature vector in each state, said mean vector varying with time in each state;

means for storing said plurality of HMMs; and

means connected to said storing means for receiving a series of unknown utterances and for employing the HMMs stored by said means for storing to determine the utterance most likely to correspond to said unknown utterance.

8. A method of speech pattern recognition employing a Hidden Markox Model (HMM), said model comprising a plurality of states, the steps comprising:

uttering a test word W for R times where R is an integer;

converting each utterance of the word W into a sequence of feature vectors;

storing each sequence of feature vectors in a memory;

determining the elements of a Markov model parameter set λ

_i utilizing said feature vector sequences, said step of determining including the determination of a neutral pint vector anda direction vector, the point vector and direction vector being related to a mean vector of a probability density function, said mean vector varying linearly with time for each state of said model;

storing said Markov model parameter set; and

employing the stored Markov model parameter set in selecting a speech pattern most likely to correspond to an unknown speech pattern.

9. A pattern recognition apparatus using a plurality of Hidden Markov Models (HMM) each having a plurality of states, comprising:

feature extraction means for converting an input signal into a sequence of feature vectors; and

HMM creating means for making an HMM from the sequence of feature vectors utilizing at least one parameter for defining a probability density function of the feature vector in each state, the probability density function comprising a mean vector varying with time in each state; and

means employing said HMM for recognizing an unknown pattern.

10. The apparatus of claim 9 wherein said mean vector varies linearly with respect to time in each state.

11. The apparatus of claim 10 wherein said probability density function further includes a term comprising a covariance matrix which is constant in time in each state.

12. The apparatus of claim 10 wherein said mean vector comprises a neutral point vector and a direction vector.

13. The apparatus of claim 9 wherein, in the course of making an HMM, said HMM creating means estimates a mean vector, said mean vector varying with time in each state.

14. The apparatus of claim 13 wherein, in the course of making an HMM, said HMM creating means further estimates a covariance matrix which is constant in each state.

15. The apparatus of claim 14 wherein said HMM creating means estimates said mean vector by deriving an estimate of a neutral point vector and a direction vector.

16. In a method of deriving a Hidden Markov Models (HMM) for speech pattern recognition, said model comprising a plurality of states, the steps comprising:

uttering a word W for R times where R is an integer;

converting each utterance of the word W into a sequence of feature vectors;

storing each sequence of feature vectors in a memory; and

calculating a Markov model parameter set λ

_i utilizing said feature vector sequences, said step of calculating including the determination of a neutral point vector and a direction vector, the neutral point vector and direction vector being related to a mean vector of a probability density function, said mean vector varying linearly with time for each state of said model.