著者

向 奈津美 金寺 登 北口 直 荒井 隆行

出版者

独立行政法人 石川工業高等専門学校

雑誌

石川工業高等専門学校紀要 (ISSN:02866110)

巻号頁・発行日

vol.39, pp.51-56, 2007

---

The process of detecting portions involving utterances, which is essential for captioning films, is generally carried out manually by translators at present. Robust methods are inevitable for automatic voice activity detection (VAD) in films involving other irrelevant sound information such as background music. This paper proposes a new feature for automatic VAD. The proposed method utilizes the gradient of spectrum in high-frequency domain (4-6kHz) and the standard deviation of modulation-filtered cepstrum. For evaluation experiments, we used a portion (about 23 minutes) of an English musical film. The proposed method exhibits a 22.6% reduction in total error rate compared to the conventional one utilizing the short time energy.