Method for reducing database requirements for speech recognition systems

• US 5,845,246 A
• Filed: 02/28/1995
• Issued: 12/01/1998
• Est. Priority Date: 02/28/1995
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: Voice Control Systems, Inc. (Koninklijke Philips N.V.)
• Inventors: Schalk, Thomas B.
• Primary Examiner(s): Hudspeth, David R.
• Assistant Examiner(s): Smits, Talivaldis Ivars
• Application Number:
  US08/396,018
• Time in Patent Office: 1,372 Days
• Field of Search: 381/42, 381/43, 395/2, 395/2.42, 395/2.52, 395/2.6, 704/233, 704/243, 704/244, 704/251
• US Class Current: 704/243
• CPC Class Codes: G10L 15/065 Adaptation
• CPC Class Codes: G10L 2015/0635 updating or merging of old and new templates; Mean values; Weighting

Abstract

The present invention comprises a method for reducing the database requirements necessary for use in speaker independent recognition systems. The method involves digital processing of a plurality of recorded utterances from a first database of digitally recorded spoken utterances. The previously recorded utterances are digitally processed to create a second database of modified utterances and then the first and second databases are combined to form an expanded database from which recognition vocabulary tables may be generated.

Claims

1. A method for reducing database requirements for a speaker independent recognizer of a speech recognition system, comprising the steps of:

selecting a plurality of utterances from a set of preexisting speech databases comprising the vocabulary of the speech recognition system;

recording into a first database samples of the selected utterances from a predetermined number of sample users;

processing each of the samples of the first database to create a plurality of modified samples of the selected utterances, the modified samples forming a second database, wherein the step of processing includes a step of filtering the samples to emphasize and de-emphasize high and low frequency information within the samples; and

combining the first database and the second database to generate an expanded database for use in training the speaker independent recognizer.

2. The method of claim 1 wherein the step of processing includes a step of applying random logarithmic scaling to the samples.

3. The method of claim 1 wherein the step of processing further includes a step of applying temporal compression and expansion to the samples to generate modified samples having altered pitch.

4. The method of claim 1 wherein the step of processing further includes a step of applying temporal compression and expansion to the samples to generate modified samples having altered format frequencies.

5. The method of claim 1 wherein the step of processing further includes a step of adding noise to each of the samples.

6. The method of claim 1 wherein the step of processing further includes a step of:

mixing samples from preexisting speech databases with samples of the first database; and

normalizing average audio levels for the mixed samples.

7. A method for reducing training database sample requirements for speaker independent recognition systems, comprising the steps of:

creating a training database of a plurality of digital samples, each digital sample representing a spoken utterance recorded from a sample user;

digitally processing each of the plurality of digital samples to generate a plurality of modified digital samples wherein the step of digitally processing includes the step of filtering the digital samples to emphasize and de-emphasize high and low frequency information within the digital samples; and

adding the plurality of modified digital samples to the training database to provide an expanded training database.

8. The method of claim 7 wherein the step of digitally processing includes a step of applying random logarithmic scaling to the digital samples.

9. The method of claim 7 wherein the step of digitally processing further includes a step of applying temporal compression and expansion to the digital samples to generate modified digital samples having an altered pitch.

10. The method of claim 7 wherein the step of digitally processing further includes a step of applying temporal compression and expansion to the digital samples to generate modified digital samples having altered format frequencies.

11. The method of claim 7 wherein the step of digitally processing further includes a step of adding digital noise to each of the digital samples to generate modified digital samples.

12. The method of claim 7 wherein the step of digitally processing further includes a step of:

mixing digital samples from a selected set of preexisting speech databases with digital samples of the first database; and

normalizing average audio levels for the mixed digital samples.