 |
International Workshop on Hands-Free Speech Communication (HSC2001)April 9-11, 2001 |
 |
|
International Workshop on Hands-Free Speech Communication (HSC2001)April 9-11, 2001 |
|
On-line garbage modeling and feature vector normalization have been successfully used to improve the noise robustness of speech recognizers. However, it is known that the balance between insertion and deletion errors is sensitive to changes in the garbage model parameter especially when feature vector normalization is applied. Previously, an SNR-based measure was used in the post-processing stage of a connected digit recognizer to reject additional insertion errors caused by the garbage model parameter change and feature vector normalization. Our objective is to improve the performance of the post-processing stage by incorporating multiple measures, in addition to the SNR, in the decision process. The proposed algorithm uses a linear combination of digit confidence, SNR, and duration. Linear Discriminant Analysis is used to combine the measures in an optimal way. The final rejection decision is made based on a single threshold. An additional advantage of the proposed approach is that new measures can be added incrementally without changing the general framework. Experimental results for German noisy connected digit recognition show that the propbsed approach achieves an average relative string-level error rate reduction of 20% over the previous SNR-based approach.
Bibliographic reference. Tang, Jian / Häkkinen, Juha / Kiss, Imre (2001): "Improved post-processing for noise robust connected digit recognition", In HSC2001, 175-158.