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

• US 20030120490A1
• Filed: 11/12/2002
• Published: 06/26/2003
• Est. Priority Date: 05/09/2000
• Status: Active Grant

• Current Assignee: Unify GmbH & Co. KG (Atos SE)
• Original Assignee: Siemens AG
• Inventors: Schneider, Tobias, Budde, Mark
• Granted Patent: US 7,844,459 B2
• Time in Patent Office: Days
• 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 a target vocabulary for training a speech recognition system, the words of the target vocabulary being converted (S8) into a phonetic description so that the individual words are represented by phonemes, and segments, composed of one or more phones, of a spoken training text which is independent of the target vocabulary being concatenated (S9) to form words of the target vocabulary in accordance with the phonetic description.

2. The method as claimed in claim 1, characterized in that segments which are as long as possible are selected and the words are concatenated from them.

3. The method as claimed in claim 2, characterized in that context information, relating to the individual segments, of one or more adjacent phones is stored (S5) in the training text and if it is not possible to concatenate all the phonemes of a word from segments with at least two phones, segments are selected which have in each case a single phone and whose phones correspond to the phonemes in the word to be concatenated which cannot be concatenated from relatively long segments and whose context information corresponds (S17) to the contexts of these phonemes in the word to be concatenated.

4. The method as claimed in claim 3, characterized in that if no segments whose context information corresponds to the contexts of the phonemes of the word to be concatenated are present when segments with individual phones are concatenated,

segments are selected which have in each case a single phone and whose phones correspond to the phonemes in the word to be concatenated which cannot be concatenated from relatively long segments and whose context information is as similar as possible (S18) to the contexts of these phonemes in the word to be concatenated.

5. The method as claimed in one of claims 1 to 4, characterized in that the segments to be concatenated are smoothed (S19) at the boundaries between two adjacent segments.

6. The method as claimed in one of claims 1 to 5, characterized in that the individual segments are energy-standardized (S19) before the concatenation.

7. The method as claimed in one of claims 1 to 6, characterized in that the segments are present in the form of electrical speech signals.

8. The method as claimed in one of claims 1 to 6, characterized in that the segments are represented by means of feature vectors.

9. The method as claimed in claim 8, characterized in that if a plurality of segments which can be assigned to the phoneme or the phonemes of the word to be concatenated are present when the segments are selected before the concatenation, a segment is determined either by averaging the feature vectors of the assignable segments, or that segment whose feature vector has the smallest average distance from the assignable segments is determined.

10. The method as claimed in one of claims 1 to 9, characterized in that the segments which are concatenated to form words are subjected to data reduction.

11. A device for creating a speech database for a target vocabulary for training a speech recognition system, with a computer system (1) which has a storage area (3) in which a computer program for executing a method as claimed in one of claims 1 to 10 is stored.