Noboru Kanedera

Relationship between speech recognition performance and related spoken document retrieval performance

Along with multimedia information including the speech information having increased, spoken document retrieval has attracted attention. Keywords or query sentences are used for spoken document retrieval. The keywords or query sentences cannot be used when we want to watch the video scenes which are related to a watching video scene. In this report, we propose a method of evaluating spoken document retrieval between video scenes, and report the results of the investigation about the relationship between the speech recognition performance and spoken document retrieval performance by using watching video scene. As a result of simulation experiment, it was confirmed to affect the search performance in order of substituted error, deletion error, insertion error, in the search both by question sentence and by the scene. The search by the scene is easy to be affected by the speech recognition error than the search by the question sentence. It was found that the influence of the insertion error in the search by t...

End‐point detection of speech using spectral transition for captioning system

In recent years, captioning video contents with text translations is increasingly necessary because of the burgeoning use of media internationally, resulting from the rapid development of communication technology. In addition, within one language, video captioning is very important for hearing‐impaired people. However, the process of captioning videos, including speech and nonspeech decisions, is often done manually by translators at present. Therefore, an efficient automatic end‐point detection of speech for captioning video contents has been proposed. We attempted to detect speech end‐points based on acoustic landmarks that identify times when acoustic changes are prominent in the speech signals [K. N. Stevens, Acoust. Phonetics (1998)]. In this study, landmarks were obtained by combining the mean square for the regression coefficients of logarithmic envelopes of 1/3‐oct bands in time, which resembles the parameter proposed by Furui to measure spectral transition [S. Furui, J. Acoust. Soc. Am. 80(4), 10...

Phoneme recognition with elliptic discrimination neural units

Many researchers achieved high phoneme recognition rates by multilayered neural networks with linear discrimination neural (LDN) units. However, it is difficult to analyze which components of the input are important to each unit in those LDN networks. This paper proposed a multilayer neural network with elliptic discrimination neural (EDN) units so that the functions of each unit in the network may be interpreted more definitely. The center of the elliptic discrimination boundary of a neural unit corresponds to a typical point in an input space. The radii of the ellipse express the extent of the corresponding components in the input space, hence it becomes clear which components of the input space are important to each unit in the EDN network. To compare the performance of EDN and LDN networks, recognition experiments of phonemes /b, d, g/ in 5240 tokens of a Japanese speech database were carried out. In the experiments, recognition rates were obtained by EDN networks as high as the rate by an LDN network. Also, it was confirmed which components of the input space are important to each unit in the EDN network.