In this paper, we present a preliminary study of spontaneous speech recognition, describing the acoustic characteristics of spontaneous speech.
Recognition of spontaneous speech is a hard problem which is known to be considerably more difficult than recognition of read speech. To get a better understanding of spontaneous speech and how it differs from read speech, a preliminary study has been carried out to compare spontaneous and read speech utterances for 4 broadcast announcers.
In spontaneous speech, the speech rate increased by a maximum of 6%. When labeling spontaneous speech, the labeling uncertainty increased by a maximum of 20% as compared to read speech. In phoneme recognition experiments we found a decrease of between 6% and 10% in the phoneme recognition accuracy for spontaneous speech.