Robust rejection for embedded systems

This paper compares three low cost rejection-modeling approaches and their combination under the framework of a voice control task. Whole word acoustic models were created with in-vocabulary training data. Rejection parameters were adjusted with validation data to find optimal rejection performance for out-of-vocabulary commands within a 3% error rate limit for in-vocabulary commands. The rejection performance was then measured with in-vocabulary and outof-vocabulary test data. The total error rate reduces from the baseline 69.4% to 10.2% for the combined approach. The impact of rejection parameters on validation data can be generalized to test data and rejection performance remains stable when only part of the vocabulary is activated during testing, indicating robustness of these approaches across different applications.