Application of inverse filtering in enhancement of whisper recognition

The differences between normal speech and whisper, particularly in terms of their acoustic characteristics, are serious problem of ASR (Automatic Speech Recognition) systems. This paper presents the preliminary results of the new way of speech signal pre-processing, which is based on inverse filtering. This method of signal pre-processing improves whisper recognition with ANNs (Artificial Neural Networks). The ANNs showed high capabilities in speech and whisper recognition in matched train/test scenarios, with the average recognition accuracy of 99.8%. However, the recognition scores in mismatched train/test scenarios were highly degraded. Because of their practical significance, the mismatched train/test scenarios were analyzed in detail in this research. Particularly, the speech/whisper scenario is important. This scenario corresponds to real life situation when speaker is in front of ASR system and from speech switches to whisper. The use of inverse filter enhanced whisper recognition by 9.48%, which in this scenario amounts 70.25%.