It is clear from different field trials that the application users may be involved in other conversations and are likely to elicit words out of the expected lexicon such as extra-words, synonyms, and extra-linguistic sounds. The user may be using the system in noisy conditions and if the system is sensitive to the environment noise it will detect a word even if the user did not speak. In order to take such phenomena into account the system has to possess rejection capabilities.
Different approaches allow implementation of such functionality. The most common approach is a rejection model which is trained on speech data from words that do not belong to the application vocabulary (called trash model ). In isolated word HMM systems, the trash model is also an HMM or several HMM models beside the models of the application vocabulary .
Another approach is usually based on thresholds about the speech recognition scores (likelihood, distance measures, etc.).
If the option of rejecting out-of-lexicon words or/and extra-linguistic phenomena is offered, then the technology provider has to explicate whether: