Method of determining an acoustic model for a word

• US 6,339,759 B1
• Filed: 09/30/1997
• Issued: 01/15/2002
• Est. Priority Date: 10/01/1996
• Status: Expired due to Fees

• Current Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Original Assignee: US Philips Corporation (Koninklijke Philips N.V.)
• Inventors: Ullrich, Meinhard D., Aubert, Xavier, Beyerlein, Peter
• Primary Examiner(s): Tsang, Fan
• Assistant Examiner(s): Opsasnick, Michael N.
• Application Number:
  US08/941,641
• Time in Patent Office: 1,568 Days
• Field of Search: 704/231, 704/240, 704/251, 704/254, 704/256
• US Class Current: 704/254
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

For the recognition of spoken text it is necessary that the words to be recognized are available in acoustically modeled form, i.e. in the form of a sequence of reference values. These reference values are determined from a known, spoken text during a training phase, in that from this text there are derived characteristic values at regular intervals, as during the recognition, which characteristic values are arranged according to triphones so as to form groups or so-called clusters. These groups constitute the basis for the reference values. In the case of a recognition system involving a very large vocabulary, however, not all triphones will occur during the training phase, unless the text is prohibitively long. In order to enable the reference values to be determined also for words containing triphones which have not occurred, such a triphone must be associated with an available group. To this end, all groups are examined so as to determine whether they have the same central phoneme in interrelationship with either the left-hand or the right-hand phoneme as the triphone to be associated. The group for which this is most often the case is selected as the associated group. The vast majority of words can thus be modeled on the basis of triphones. Modifications of this rule are described for words containing triphones which cannot be directly associated in this manner.

Claims

1. A method of determining an acoustic model for a speech recognition system comprising the steps of:

(i) during a training phase, deriving characteristic values of triphones in a speech test signal, which characteristic values represent acoustic triphone states, each triphone consisting of a central phoneme, a left-hand phoneme and a right-hand phoneme;

(ii) combining the characteristic values of triphones which satisfy predetermined criteria into respective groups of triphones;

(iii) for modeling of a triphone which has not been observed during said training phase, selecting for said unobserved triphone one of the groups of triphones having a same central phoneme and either a same left-hand phoneme or right-hand phoneme.

2. A method as claimed in claim 1, wherein for each acoustic state in a triphone to be modeled there is selected that group which is associated with such acoustic state of the observed triphone having the largest number of combinations of the same central phoneme and either the same left-hand or right-hand phoneme as the triphone to be modeled.

3. A method as claimed in claim 2, wherein for the first acoustic state in the triphone to be modeled the number of combinations with the same central phoneme and the same left-hand phoneme is increased by a fixed value, and for the last acoustic states of said triphone the number of combinations with the same central phoning and the same right-hand phoneme is increased by a fixed value.

4. A method as claimed in claim 1, wherein for each acoustic state in the triphone to be modeled a group is searched for which is associated with that acoustic state in an observed triphone having the same central phoneme and the same left-hand or right-hand phoneme, and if such a group is not available then a group is searched for which is associated with a neighboring acoustic state.

5. A method as claimed in claim 1, wherein in the case of absence of a group associated with observed triphones having the same central phoneme and the same left-hand or right-hand phoneme as the triphone to be modeled, a group is searched for which is associated with the same left-hand or right-hand phoneme but other central phonemes, the number of such triphones in a group which is examined being weighted.