ASR Failure Prediction Based on Signal Measures

Authors: Lu Huo, Ulrich Heute

Abstract:

In a spoken-dialogue system, it is useful to know to which extent the automatic speech recognizer (ASR) can overcome the adversary environment of a real-world telephone system and still produce correct recognition results. This kind of information, from the system developer’s point of view, can be used to adjust the recognition-result handling; and from the system-manager’s point of view, can be used as an important system-evaluation factor. In this work, we study signal-based measures that influence the recognition performance, and we try to build a prediction model for the “Lucy” database.