Method for speech recognition using partitioned vocabulary

• US 8,306,820 B2
• Filed: 10/04/2005
• Issued: 11/06/2012
• Est. Priority Date: 11/16/2004
• Status: Expired due to Fees

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Kunstmann, Niels
• Primary Examiner(s): Smits, Talivaldis Ivars
• Assistant Examiner(s): ROBERTS, SHAUN A
• Application Number:
  US11/667,825
• Publication Number:
  US 20080126090A1
• Time in Patent Office: 2,590 Days
• Field of Search: 704/245, 704/251
• US Class Current: 704/245
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 2015/085 Methods for reducing search complexity, pruning

Abstract

A is recognized using a predefinable vocabulary that is partitioned in sections of phonetically similar words. In a recognition process, first oral input is associated with one of the sections, then the oral input is determined from the vocabulary of the associated section.

Claims

1. A method for recognizing a spoken input using a predefinable vocabulary, comprising:

organizing the predefinable vocabulary, prior to receiving the spoken input, based on distance measures of phonetic similarity between pairs of words in the predefinable vocabulary by

obtaining ranking values of each pair of words in the predefinable vocabulary as possible matches for test utterances independent of the pair of words,

averaging, for each pair of words in the predefinable vocabulary, differences between the ranking values for all of the test utterances, to obtain the distance measures;

storing, on a storage medium, the predefinable vocabulary subdivided into sections of phonetically similar words based on the distance measures and using a vector quantization algorithm; and

characterizing each of the sections of the phonetically similar words by a representative entry;

assigning the spoken input to one of the sections for which the representative entry is most similar to the spoken input; and

identifying a closest match for the spoken input among the phonetically similar words in the one of the sections that has been assigned to the spoken input.

2. The method as claimed in claim 1, wherein the spoken input is at least one spoken word.

3. The method as claimed in claim 1, wherein, said determining of the distance measure for phonetic similarity between the pairs of words comprises:

determining distance values for a similarity of two letter sequences; and

adding the distance values for the distance measure of the two letter sequences.

4. The method as claimed in claim 3, wherein said storing stores the sections of the predefinable vocabulary by different lengths of letter sequences.

5. The method as claimed in claim 4, wherein a Levenshtein distance is used as the distance measure.

6. A computer readable medium encoding a computer program which when executed by a processor causes the processor to perform a method comprising:

organizing the predefinable vocabulary, prior to receiving the spoken input, based on a distance measure for phonetic similarity between pairs of words in the predefinable vocabulary by

obtaining ranking values of each pair of words in the predefinable vocabulary as possible matches for test utterances independent of the pair of words, and

averaging, for each pair of words in the predefinable vocabulary, differences between the ranking values of the pairs of words for the test utterances to obtain the distance measures;

storing, on a storage medium, the predefinable vocabulary subdivided into sections of phonetically similar words based on the distance measures and using a vector quantization algorithm; and

characterizing each of the sections of the phonetically similar words by a representative entry;

assigning the spoken input to one of the sections for which the representative entry is most similar to the spoken input; and

identifying a closest match for the spoken input among the phonetically similar words in the one of the sections that has been assigned to the spoken input.