Confidence measures using sub-word-dependent weighting of sub-word confidence scores for robust speech recognition

• US 6,539,353 B1
• Filed: 10/12/1999
• Issued: 03/25/2003
• Est. Priority Date: 10/12/1999
• Status: Expired due to Term

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Jiang, Li, Huang, Xuedong
• Primary Examiner(s): SMITS, TALIVALDIS IVARS
• Application Number:
  US09/417,173
• Time in Patent Office: 1,260 Days
• Field of Search: 704/252, 704/254, 704/255
• US Class Current: 704/254
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A method and apparatus is provided for speech recognition. The method and apparatus convert an analog speech signal into a digital signal and extract at least one feature from the digital signal. A hypothesis word string that consists of sub-word units is identified from the extracted feature. For each identified word, a word confidence measure is determined based on weighted confidence measure scores for each sub-word unit in the word. The weighted confidence measure scores are created by applying different weights to confidence scores associated with different sub-words of the hypothesis word.

Claims

1. A speech recognition system for identifying words from an input digital signal representing speech, the system comprising:

a feature extractor for extracting at least one feature from the input digital signal;

a search engine capable of identifying at least one hypothesis word and at least one speech recognizer score for a sub-word of the hypothesis word based on at least one feature and at least one speech model; and

a confidence measure module capable of determining a confidence measure based in part on a confidence score for a sub-word of the hypothesis word, the confidence measure attributing different weights to confidences scores of different classes of sub-words in the hypothesis word.

2. The speech recognition system of claim 1 wherein attributing different weights to different sub-words comprises attributing different parameters to functions of a sub-word confidence score.

3. The speech recognition system of claim 2 wherein the functions are linear functions.

4. The speech recognition system of claim 2 wherein the functions are quadratic functions.

5. The speech recognition system of claim 2 further comprising a trainer for training the parameters of the functions.

6. The speech recognition system of claim 5 wherein the trainer uses a gradient descent algorithm to train the parameters.

7. The speech recognition system of claim 6 wherein the gradient descent algorithm is based on a loss function.

8. The speech recognition system of claim 6 wherein sub-words are assigned to classes such that sub-words within a class are weighted identically and wherein the gradient descent algorithm comprises updating each parameter of each class based on a respective gradient associated with each parameter.

9. A method of speech recognition comprising:

extracting at least one feature from a set of digital values that represent a speech signal;

identifying a hypothesis word formed of sub-words from the feature;

determining sub-word confidence scores from the feature for each sub-word of the hypothesis word; and

determining a word confidence score for the hypothesis word by-applying different weights to the sub-word confidence scores associated with different classes of sub-words of the hypothesis word.

10. The method of claim 9 wherein applying different weights to sub-word confidence scores comprises applying different parameters to functions of the sub-word confidence scores.

11. The method of claim 10 wherein applying different weights to sub-word confidence scores further comprises selecting the parameters for a class based on discriminative training.

12. The method of claim 11 wherein discriminative training comprises determining a gradient of a loss function for each parameter associated with a word.

13. The method of claim 12 wherein discriminative training further comprises updating each parameter on a pass through a training text, each parameter updated based on an average gradient of the loss function for the respective parameter across the training text.

14. The method of claim 9 wherein determining a sub-word confidence score comprises identifying a recognition acoustic model score for a sub-word and determining the ratio of the recognition acoustic model score over a secondary acoustic model score for the sub-word.

15. The method of claim 14 wherein determining a word confidence score further comprises weighting multiple confidence scores for each sub-word.

16. A computer-readable medium having computer executable instructions for performing steps comprising:

receiving a digital signal representative of an input speech and noise signal;

extracting features from the digital signal;

identifying a recognition score from the features for a hypothesis word formed of sub-words; and

determining a confidence score for the hypothesis word based on weighted sub-word confidence scores determined from the features, the weighting of sub-word confidence scores applying different weights to sub-word confidence scores associated with different classes of sub-words for the sub-words of the word.

17. The computer-readable medium of claim 16 wherein applying different weights to sub-word confidence scores comprises applying different parameters to functions of the sub-word confidence scores.

18. The computer-readable medium of claim 17 wherein applying different weights to sub-word confidence scores further comprises associating each sub-word with a class such that the functions of the sub-word confidence scores associated with sub-words within a class all have the same parameters.