A Study of Paralinguistic Features Natural Japanese Dialogue Speech
2001 Satoshi Kobayashi
Spoken language has two features those are not segmental, prosody and paralanguage. In the first stage, I compared prosodic features and paralinguistic features. Prosodic features represent distinction of meanings, while paralinguistic features represent intentions, attitudes, feelings and others.
The existence of prosodic features can be judged objectively based on linguistic features, like meanings. On the other hand, the existence of paralinguistic features can be judged based on only subjective feeling. However, if judgment of the existence of paralinguistic features is unstable, it is difficult to conduct research about paralinguistic features. Therefore, subjects attached labels of paralinguistic features in terms of height, loudness, tempo and emphasis of spontaneous speech to assess the stability of transcribers' judgments.
Next, the degree of paralanguage transmission was evaluated by investigating how many of the same labels ware put at the same locations in the utterances. As a result, 40Vo of.labels corresponded among transcribers. However, I0Vo of labels conflicted among some of the transcribers, and 50Vo of labels were written down by only one transcriber. This means that most of the differences of labels among transcribers are caused by oversight, that is, by the difference of significance level of individuals.
Furthermore, the subjects agreed with each other regarding the existence or lack of labels. This result shows that the subjects perceived paralinguistic features in the same manner.
Next, I investigated the correlation between paralinguistic features and prosody. The results show that for a small number
of labels it is difficult to estimate whether they belong to a paralinguistic feature or prosody, and it is assumed that the paralinguistic features are perceived in distinction from prosody.
Next, I discussed the perceptual differences :rmong the three paralinguistic features.
I also discussed acorstic parameters at the labeled locations. It is important to note that 80Vo of the labels corresponded to physical changes. Larger changes showed a higher possibility of being noticed and labeled by transcribers, and also showed a higher possibility of being labeled corresponding to physical changes. However, when a change was relatively large, the highest rate of labeling for each target reached 307o. This means that paralinguistic features are not necessarily judged only based on physical features.
Lastly, I compared labels for utterances with phonetic content and synthesized utterances without phonetic content, and tested the effect of the content of speech sounds. As a result, loudness was affected by phonetic content.
ln
‑132‑
