Phone speech detection and recognition in the task of historical radio broadcast transcription

This paper deals with methods and strategies for the improvement of a system for the automatic transcription of the historical Czech Radio audio archive. The main goal of this work was to improve the recognition of audio signals containing phone speech where the resulting recognition rate was relatively low because of frequency-limited phone signals. A phone signal detector based on GMM was developed and implemented to our speech transcription system. Several different acoustic models were experimentally tested for the enhancement of phone speech signal recognition. We demonstrate that phone speech recognition is improved significantly if acoustic models based on HMM are trained directly on phone speech signals. Other possible and logical strategies, which are described in this paper, did not produce the required improvement. The resulting accuracy of phone speech signal recognition has been increased from 47.32% to 68.30%.