Research On Confidence Measure In Speech Keyword Spotting

As an important technology of speech recognition, keyword spotting has broad application space and great practical value. Although there are a lot of researches about the keyword spotting technology, errors are still inevitable in the recognition results, due to the influence of adverse factors, such as environmental noise, pronunciation differences of speakers and so on. Confidence measure can be used to evaluate the reliability of the candidate keywords objectively without reference alignment. Therefore it is crucial for the keyword spotting. According to the different properties of hypothesized words, the methods of confidence measure and normalization for In-Vocabulary (IV) and Out-of-Vocabulary (OOV) words are studied respectively in this paper. Research contributions of this thesis are as follows:(1) To solve the problem of insufficient use of contextual information, a confidence measure method for In-Vocabulary words based on optimization of context consistency is presented. Firstly, the reliability of the hypothesized words is evaluated by the context consistency. Then an adaptive sliding window is used to segment the semantic fragments of the speech recognition results to remove consistency interference, after which the context consistency of hypothesized words can be forced to calculate just in its own semantic fragment. Furthermore, the distance information between words is introduced to calculate context consistency to distinguish the influence of diverse position context words. Experimental results show that the proposed optimal measures could improve the accuracy of context consistency greatly.(2) As some In-Vocabulary words may be recognized as Out-of-Vocabulary words in the soft match-based OOV detection, we propose a confidence measure method for OOV based on error types in the speech recognition. The errors of recognition are estimated according to the properties of hypothesized words. Then the recognition error types are turned into features and a distinguish model is trained to detect the IV/OOV positions. If the OOV candidate based on soft match is found in the IV area, it can be treated as a false alarm. Otherwise the OOV candidate is in the OOV area, a fused confidence measure will be employed to calculate the reliability of the hypothesized words. Experimental results show that the confidence measure based on extended errors can improve the performance of OOV words detection greatly.(3) As different hypothesized words have different properties, the same confidence scores may represent different confidence level. A term-dependent confidence normalization method based on ATWV (Actual Term-Weighted Value) optimization is proposed. Firstly, we adjust the words’confidence score according to their frequency in the test. Then the confidence bias in ATWV optimization is compensated in a linear and discriminative way respectively. The linear compensation adjusts confidence by adding weighted and translation factors. While the discriminative compensation convert confidence score to classification posterior probability, which meets the requirements of ATWV optimization, by discriminative model training. Experimental results show that the proposed confidence normalization is effective for the keyword spotting.