Voice recognition system and voice recognition method

A voice recognition system, comprising:

• voice segmentation unit which calculates a voice feature value based on time-series input sound data, segmenting one of a section of voice of said time-series input sound data and a section of non-voice of said time-series input sound data by comparing a threshold value and said voice feature value, and determining one of said section and a section made by adding a margin of designated length before and after said original section as a first voice section;
• search unit which determines a section to be a target of a voice recognition as a second voice section based on a likelihood of voice and a likelihood of non-voice calculated based on a voice recognition feature value;
• parameter updating unit which updates at least one of said threshold value and said margin used upon said voice segmentation unit determining the first voice section according to a difference between length of said first voice section and length of said corresponding second voice section;
• and
• said voice segmentation unit which determines the first voice section using one of the threshold value and the margin updated by said parameter updating unit.

A voice recognition method:

• calculates a voice feature value based on time-series input sound data;
• segments one of a section of voice of said time-series input sound data and a section of non-voice of said time-series input sound data by comparing a threshold value and said voice feature value;
• determines one of said section and a section made by adding a margin of designated length before and after said original section as a first voice section;
• determines a section to be a target of a voice recognition as a second voice section based on a likelihood of voice and a likelihood of non-voice calculated based on a voice recognition feature value;
• updates at least one of said threshold value and said margin used upon determining said first voice section according to a difference between length of said first voice section and length of said corresponding second voice section;
• and,
• upon segmenting said first voice section, determines a first voice section using one of the updated threshold value and the margin.

A non-transitory computer-readable program recording medium for storing a voice recognition program causing

• a computer to execute:
• voice segmenting processing for calculating a voice feature value based on time-series input sound data, segmenting one of a section of voice of said time-series input sound data and a section of non-voice of said time-series input sound data by comparing a threshold value and said voice feature value, and determining one of said section and a section made by adding a margin of designated length before and after said section as a first voice section;
• search processing for determining a section to be a target of a voice recognition as a second voice section based on a likelihood of voice and a likelihood of non-voice calculated based on a voice recognition feature value;
• parameter updating processing for updating at least one of said threshold value and said margin used upon determining said first voice section according to a difference between length of said first voice section and length of said corresponding second voice section;
• and,
• in said voice segmenting processing, making the first voice section be determined using one of the threshold value and the margin updated in said parameter updating processing.