Kokusai Denshin Denwa Co., Ltd.

Nagoya University

Equos Research Co., Ltd.

Kazuya Takeda

Shingo Kuroiwa

Seiichi Yamamoto

Seiichi Suzuki

Hirokazu Nomoto

Masaki Naito

Fumihiro Yato

Kazuaki Fujii

Katunobu Itou

Chiyomi Miyajima

Koji Ozawa

A driver model with higher precision is created as an evaluation standard for a driving condition in a normal condition. By detecting biometric information of a driver, whether a driver is in a usual condition or not is recognized. Then, data of driving conditions are collected while the driver is driving, and from the driving condition data, a part indicating that the driver operates in a usual condition is extracted to create a driver model. Further, the driver model is created taking only a case of driving in a normal condition as a driving action in normal times based on biometric information of the driver, and hence the driver model becomes more precise and neutral. Further, by using a GMM (Gaussian mixture model) for the driver model, a driver model for each driver can be created easily, and moreover, by calculation to maximize a conditional probability, a driving operation action is easily estimated and outputted.

Driving action estimating device, driving support device, vehicle evaluating system, driver model creating device, and driving action determining device

A speech endpoint detection unit determines that a first condition is met when a matching score of a partial sentence accepted by a grammatical rule unit is the highest of all partial sentences, and that a second condition is met when a duration time of input speech determined to coincide with a silent standard pattern is longer than a predetermined time. A speech endpoint is determined when the first and second conditions are both met and a speech endpoint detection signal is sent to a word prediction unit and a recognition result output unit. By requiring that both first and second conditions be met, a speech endpoint can correctly be detected even when a long silent period is present in the course of a sentence.

Speech endpoint detection method and apparatus and continuous speech

This invention relates to a method of recognizing input speech of many unspecific people. Feature parameters representing both a short-time average spectrum envelope characteristic of the input speech, and regression coefficients obtained from the outputs from a wide-band filter bank. The regression coefficients represent the rough directionality of the characteristic of change in the spectrum of the speech signal. Distance is measured between the feature parameters and standard patterns stored in a storage means. The distance between the feature parameters and the stored pattern which is smallest of all the patterns is found to recognize said input speech.

Speech recognition method

Growing community of inventors

Nagoya, Japan

Kazuya Takeda