In this paper we investigate the exploitation of loosely transcribed audio data, in the form of captions for weather forecast recordings, in order to adapt acoustic models for automatically transcribing these kinds of forecasts. We focus on dealing with inaccurate time stamps in the captions and the fact that they often deviate from the exact spoken word sequence in the forecasts. Furthermore, different adaptation algorithms are compared when incrementally increasing the amount of adaptation material, for example, by recording new forecasts on a daily basis.

Index Terms: speech recognition, acoustic model adaptation, slightly supervised training, loose transcripts, adaptation methods