A System for Estimating the Importance of Speech Based on Acoustic Features

With the development of AI technology, the accuracy of speech recognition and the range of its use are advancing. With AI-based speech processing, many services such as automatic speech transcription, speech recognition, and speech summarization are now available. In this paper, a method is proposed for determining whether each utterance is important or not based on acoustic information of the utterances. As acoustic features, statistical measures of various acoustic features for each utterance are used. To determine the importance of the utterance, the importance of the transcribed text is labeled using the LLM chat system and trained as a supervised input. In this experiment, TED video speeches on YouTube are used as speech materials. A machine learning model with a random forest classifier is used to determine the importance. As a result, a model that can classify the importance for the training data is obtained. It is found that statistical measures of acoustic features related to the fundamental frequency (F0) of the utterance are frequently used as important features for classification. However, when evaluated on test data, it is found that sufficient accuracy is not achieved, and further examination is necessary.