Method for creating a speech database for a target vocabulary in order to train a speech recognition system

• US 7,844,459 B2
• Filed: 04/24/2001
• Issued: 11/30/2010
• Est. Priority Date: 05/09/2000
• Status: Active Grant

• Current Assignee: Unify GmbH & Co. KG (Atos SE)
• Original Assignee: Siemens AG
• Inventors: Budde, Mark, Schneider, Tobias
• Primary Examiner(s): Han, Qi
• Application Number:
  US10/276,071
• Publication Number:
  US 20030120490A1
• Time in Patent Office: 3,507 Days
• Field of Search: 704/254, 704/255, 704231-245, 704/251
• US Class Current: 704/254
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

The words of the target vocabulary are composed of segments, which have one or more phonemes, whereby the segments are derived from a training text that is independent from the target vocabulary. The training text can be an arbitrary generic text.

Claims

1. A method for creating a speech database for training of a speech recognition system with a target vocabulary, comprising:

converting the words of the target vocabulary into a phonetic description so that the individual words are represented by a sequence of phonemes; and

concatenating, using a computer, segments of a generic spoken training text to form the words of the target vocabulary and thereby the speech database, each segment containing at least one phone, the segments being concatenated to form a sequence of phones which correspond respectively to the sequence of phonemes of the target vocabulary.

2. The method as claimed in claim 1, wherein the segments of the generic spoken training text are selected for concatenation such that segments having more phones are selected over segments having less phones and the selected segments are as long as possible.

3. The method as claimed in claim 2, wherein

for at least segments having only one phone, context information, relating to phones of adjacent segments phones is stored, and

if it is not possible to form a word of the target vocabulary by concatenating only segments with at least two phones, a segment having only one phone is selected if:

the phone of the segment corresponds to phoneme in the word of the target vocabulary, and

the context information for the segment indicates that adjacent phones correspond to phonemes adjacent to the phoneme in the word of the target vocabulary.

4. The method as claimed in claim 3 wherein if no segment of the generic spoken training text has two adjacent phones corresponding to two phonemes adjacent to the phoneme in the word of the target vocabulary, a segment is selected if:

the phone of the segment corresponds to the phoneme in the word of the target vocabulary, and

the context information for the segment indicates that one of the two adjacent phones corresponds to the respective adjacent phoneme in the word of the target vocabulary.

5. The method as claimed in claim 4, wherein the segments to be concatenated are smoothed at the boundaries between two adjacent segments.

6. The method as claimed in claim 5, wherein the individual segments are energy-standardized before the concatenation.

7. The method as claimed in claim 6, wherein the segments are in the form of electrical speech signals.

8. The method as claimed in claim 6, wherein the segments are represented by feature vectors.

9. The method as claimed in claim 8, wherein if a plurality of potential segments are available for concatenation, a calculated segment is selected for concatenation, the calculated segment being produced by averaging feature vectors associated with each of the potential segments.

10. The method as claimed in claim 9, wherein the phones of the generic spoken training text are saved in a format which reduces data.

11. The method as claimed in claim 8, wherein if a plurality of potential segments are available for concatenation, a calculated segment is selected for concatenation, the calculated segment being produced by obtaining a segment whose feature vector has the smallest average distance from feature vectors of the potential segments.

12. The method as claimed in claim 11, wherein the phones of the generic spoken training text are saved in a format which reduces data.

13. The method as claimed in claim 1, wherein the segments to be concatenated are smoothed at the boundaries between two adjacent segments.

14. The method as claimed in claim 1, wherein the individual segments are energy-standardized before the concatenation.

15. The method as claimed in claim 1, wherein the segments are in the form of electrical speech signals.

16. The method as claimed in claim 1, wherein the segments are represented by feature vectors.

17. The method as claimed in claim 16, wherein if a plurality of potential segments are available for concatenation, a calculated segment is selected for concatenation, the calculated segment being produced by averaging feature vectors associated with each of the potential segments.

18. The method as claimed in claim 16, wherein if a plurality of potential segments are available for concatenation, a calculated segment is selected for concatenation, the calculated segment being produced by obtaining a segment whose feature vector has the smallest average distance from feature vectors of the potential segments.

19. The method as claimed in claim 1, wherein the phones of the generic spoken training text are saved in a format which reduces data.

20. A computer readable medium storing a program to control a computer to perform a method for creating a speech database for training of a speech recognition system with a target vocabulary, said method comprising:

converting the words of the target vocabulary into a phonetic description so that the individual words are represented by a sequence of phonemes, and

concatenating segments of a generic spoken training text to form the words of the target vocabulary and thereby the speech database, each segment containing at least one phone, the segments being concatenated to form a sequence of phones which correspond respectively to the sequence of phonemes of the target vocabulary.

21. A method for creating a speech database for training of a speech recognition system with a target vocabulary, comprising:

converting the words of the target vocabulary into a phonetic description so that the individual words are represented by a sequence of phonemes;

obtaining a generic spoken training text produced from speech of at least 100 speakers; and

concatenating, using a computer, segments of the generic spoken training text to form the words of the target vocabulary and thereby the speech database, each segment containing at least one phone, the segments being concatenated to form a sequence of phones which correspond respectively to the sequence of phonemes of the target vocabulary,

wherein the speech database is created without requiring speakers to speak the entire target vocabulary.