Method of generating a subword model for speech recognition

• US 5,677,988 A
• Filed: 09/21/1995
• Issued: 10/14/1997
• Est. Priority Date: 03/21/1992
• Status: Expired due to Fees

• Current Assignee: ATR Interpreting Telephony Research Laboratories
• Original Assignee: ATR Interpreting Telephony Research Laboratories
• Inventors: Takami, Jun-ichi, Sagayama, Shigeki
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Chawan, Vijay B.
• Application Number:
  US08/532,318
• Time in Patent Office: 754 Days
• Field of Search: 395/2.4, 395/2.47, 395/2.52, 395/2.58, 395/2.59, 395/2.65, 395/2.75
• US Class Current: 704/256
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/146 with insufficient amount of training data, e.g. state sharing, tying, deleted interpolation
• CPC Class Codes: G10L 2015/022 Demisyllables, biphones or triphones being the recognition units

Abstract

An automated method of generating a subword model for speech recognition dependent on phoneme context for processing speech information using a Hidden Markov Model in which static features of speech and dynamic features of speech are modeled as a chain of a plurality of output probability density distributions. The method comprising determining a phoneme context class which is a model unit allocated to each model, the number of states used for representing each model, relationship of sharing of states among a plurality of models, and output probability density distribution of each model, by repeating splitting of a small number of states, provided in an initial Hidden Markov Model, based on a prescribed criterion on a probabilistic model.

Claims

1. An automated method of generating a subword model for speech recognition dependent on phoneme context for processing speech information using a Hidden Markov Model in which static features of speech and dynamic features of speech are modeled as a chain of a plurality of output probability density distributions, comprising the step of:

determining a phoneme context class which is a model unit allocated to each model, the number of states used for representing each model, relationship of sharing of states among a plurality of models, and output probability density distribution of each model, by repeating splitting of a small number of states, provided in an initial Hidden Markov Model.

2. The method according to claim 1, wherein splitting includes

the step of splitting, in a parallel domain, one state into two states corresponding to different phoneme context classes to absorb change in the static features of speech derived from a difference of the phoneme context, and the step of splitting, in a serial domain, one state into two states corresponding to different phoneme segments to absorb change in the dynamic features of speech generated in a certain phoneme context class, and

repeating state splitting in the domain which enables a higher value of an evaluation with respect to an actual speech sample to successively make the model structure more precise.

3. The method according to claim 2, further comprising the step of:

in state splitting in a contextual domain, splitting with respect to actual phoneme context class and allocating two split phonetic context classes, to attain a highest value of an evaluation with respect to the actual speech sample, to one of the two states respectively generated by state splitting to make the model unit successively smaller.

4. The method according to claim 1, further comprising the step of:

allocating a mixture gaussian density distribution having the mixture number of 2 as each output probability density distribution, and

allocating a single gaussian distribution of one of two distributions constituting the mixture gaussian density distribution to two states newly generating by the state splitting, to realize significant reduction in an amount of calculation necessary for re-estimating output probability density distribution parameters after state splitting.

5. The method according to claim 4, further comprising the step of:

every time state splitting is executed, recovering the single gaussian distribution generated by state splitting to the original mixture gaussian density distribution having the mixture number of 2, and carrying out re-training to optimize model parameters.

6. The method according to claim 1, further comprising the step of:

prior to state splitting, calculating a magnitude of all states existing in a speech parameter space by determining a state having a largest magnitude as a splittee state, thereby significantly reducing an amount of calculation by avoiding a round robin type calculation necessary for determining an optimal splittee state.

7. The method according to claim 4, further comprising the step of:

after state splitting is completed and a final model unit and structure are determined, carrying out re-training to replace the mixture gaussian density distribution having the mixture number of 2, allocated to each state, by the output probability density distribution to be used in an actual Hidden Markov Network.