Joint discriminative training of multiple speech recognizers
View Patent ↗Adjusting model parameters is described for a speech recognition system that combines recognition outputs from multiple speech recognition processes. Discriminative adjustments are made to model parameters of at least one acoustic model based on a joint discriminative criterion over multiple complementary acoustic models to lower recognition word error rate in the system.
1. A method of adjusting model parameters in a speech recognition system comprising:
in a speech recognition system executed by computer system that combines recognition outputs from a plurality of parallel speech recognition processes that operate on a given speech input word sequence and produce different competing recognition outputs which are then combined to determine a final recognition output;
wherein the speech recognition processes are complementary so as to produce different recognition errors on a given same speech input;
performing a discriminative adjustment process including:
i. selecting at least one acoustic model of the system, and
ii. adjusting a plurality of model parameters of the selected acoustic model based on a joint discriminative criterion over a plurality of complementary acoustic models to lower combined recognition WER in the system over only independently discriminatively trained recognition systems.
2. A method according to claim 1 , wherein the complementary acoustic models include the at least one acoustic model.
3. A method according to claim 1 , wherein the speech recognition system combines recognition outputs based on a Confusion Network Combination (CNC) approach.
4. A method according to claim 1 , wherein the speech recognition system combines recognition outputs based on a Recognizer Output Voting for Error Reduction (ROVER) approach.
5. A method according to claim 1 , wherein the joint criterion is based on Minimum Classification Error (MCE) training.
6. A method according to claim 1 , wherein the joint criterion is based on Maximum Mutual Information Estimation (MMIE) training.
7. A method according to claim 1 , wherein adjusting a plurality of model parameters includes using a Gradient Descent algorithm.
8. A method according to claim 1 , wherein adjusting a plurality of model parameters includes using an Extended Baum-Welch algorithm.
9. A method according to claim 1 , wherein the lower recognition word error rate reaches at least a 5% word error rate reduction.
10. A method according to claim 1 , wherein discriminative adjusting a plurality of model parameters includes performing parameter adaptation using at least one linear transformation.
11. A method according to claim 1 , wherein discriminative adjusting a plurality of model parameters includes performing parameter adaptation using at least one non-linear transformation.
12. A speech recognition system containing acoustic models having model parameters adjusted by the method according to any of claims 1 - 11 .