On the integration of dialect and speaker adaptation in a multi-dialect speech recognition system

The recognition accuracy in recent Automatic Speech Recognition (ASR) systems has proven to be highly related to the correlation of the training and testing conditions. Several adaptation approaches have been proposed in an effort to improve the speech recognition performance, and have typically been applied to the speaker- and channel-adaptation tasks. We have shown in the past that a mismatch in dialects between the training and testing speakers significantly influences the recognition accuracy, and we have used adaptation to compensate for this mismatch. The dialect of the speaker needs to be identified in a dialect-specific system, and in this paper we present results in this area. To achieve further improvement in recognition performance, we combine dialect- and speaker-adaptation.