Error-Responsive Modifications to Speech Recognizers: Negative N-grams

We describe an error analysis technique that facilitates blame assignment among the various components of a speech recognizer and provides insight into their behavior. Tools are presented that help clarify how each of the component models and their interactions contribute to the bottom line performance. We use this technique to study the performance of the backoff [4] language model. The analysis highlights the significant effect of negative n-grams - sequences of words not seen in the training data. This leads to two modifications to the decoder, both of which are presented with experimental results. The first modification failed so far to improve recognition performance. The second yields up to 4% reduction in word error rate.