Frequent word section extraction in a presentation speech by an effective dynamic programming algorithm

Word frequency in a document has often been utilized in text searching and summarization. Similarly, identifying frequent words or phrases in a speech data set for searching and summarization would also be meaningful. However, obtaining word frequency in a speech data set is difficult, because frequent words are often special terms in the speech and cannot be recognized by a general speech recognizer. This paper proposes another approach that is effective for automatic extraction of such frequent word sections in a speech data set. The proposed method is applicable to any domain of monologue speech, because no language models or specific terms are required in advance. The extracted sections can be regarded as speech labels of some kind or a digest of the speech presentation. The frequent word sections are determined by detecting similar sections, which are sections of audio data that represent the same word or phrase. The similar sections are detected by an efficient algorithm, called Shift Continuous Dynamic Programming (Shift CDP), which realizes fast matching between arbitrary sections in the reference speech pattern and those in the input speech, and enables frame-synchronous extraction of similar sections. In experiments, the algorithm is applied to extract the repeated sections in oral presentation speeches recorded in academic conferences in Japan. The results show that Shift CDP successfully detects similar sections and identifies the frequent word sections in individual presentation speeches, without prior domain knowledge, such as language models and terms.