Method and apparatus for predicting word error rates from text

• US 7,117,153 B2
• Filed: 02/13/2003
• Issued: 10/03/2006
• Est. Priority Date: 02/13/2003
• Status: Expired due to Fees

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Chelba, Ciprian, Acero, Alejandro, Gunawardana, Asela J. R., Mahajan, Milind, Deng, Yonggang
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US10/365,850
• Publication Number:
  US 20040162730A1
• Time in Patent Office: 1,328 Days
• Field of Search: 704/236, 704/240, 704/241, 704/242, 704/243, 704/244, 704/255, 704/256
• US Class Current: 704/236
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/197 Probabilistic grammars, e.g. word n-grams

Abstract

A method of modeling a speech recognition system includes decoding a speech signal produced from a training text to produce a sequence of predicted speech units. The training text comprises a sequence of actual speech units that is used with the sequence of predicted speech units to form a confusion model. In further embodiments, the confusion model is used to decode a text to identify an error rate that would be expected if the speech recognition system decoded speech based on the text.

Claims

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

decoding a test text comprising actual sequences of speech units to produce sequences of predicted speech units using a confusion model that provides likelihoods for sequences of predicted speech units given sequences of actual speech units, wherein the confusion model comprises a set of rules, each rule providing a conversion from an actual speech unit into at least one predicted speech unit and a probability of performing such a conversion; and

determining an error rate based on a sequence of predicted speech units and the sequence of actual speech units.

2. The computer-readable medium of claim 1 wherein at least one rule is context independent.

3. The computer-readable medium of claim 1 wherein at least one rule is context dependent.

4. The computer-readable medium of claim 3 wherein a first rule for converting an actual speech unit into at least one predicted speech unit is context dependent on a first number of actual speech units and a second rule for converting the actual speech unit into the at least one predicted speech unit is context dependent on a second number of actual speech units.

5. The computer-readable medium of claim 1 having computer-executable instructions for performing further steps comprising training the confusion model.

6. The computer-readable medium of claim 5 wherein training the confusion model comprises:

decoding a speech signal produced from a training text to produce a sequence of predicted speech units;

identifying a sequence of training speech units in the training text;

aligning the sequence of predicted speech units with the sequence of training speech units; and

using the aligned sequences to train the confusion model.

7. The computer-readable medium of claim 6 wherein using the aligned sequences to train the confusion model comprises identifying the predicted speech units that a training speech unit is aligned with in the aligned sequences.

8. The computer-readable medium of claim 7 wherein training the confusion model further comprises generating a rule for each unique combination of predicted speech units aligned with a particular training speech unit.

9. The computer-readable medium of claim 8 wherein training the confusion model further comprises determining a probability for each rule by counting the number of times the combination of predicted speech units and training speech unit associated with the rule are aligned in the aligned sequences.

10. The computer-readable medium of claim 9 wherein training the confusion model further comprises pruning a rule if the combination of predicted speech units and training speech unit associated with the rule are aligned less than a threshold number of times.