Subtitle generating apparatus, subtitle generating method, and storage medium

A subtitle generating method executed under control of a processing circuit of a subtitle generating apparatus, the subtitle generating apparatus including the processing circuit and a display, and the method comprising:

• sequentially acquiring texts from voice recognition results;
• storing the texts as archive datasets;
• estimating a split position and a concatenation position of the texts from one or more of the archive datasets;
• generating a subtitle text from said one or more of the archive datasets based on the split position and the concatenation position;
• updating the archive datasets based on the split position and the concatenation position;
• and
• displaying the subtitle text on the display,
• wherein:
• the estimating comprises:
• continuing estimation of the concatenation position so as to concatenate the texts in the archive datasets, until a number of characters of an obtained concatenated text reaches a threshold character number or greater,
• subjecting the concatenated text to one or both of a morphological analysis and a dependency parsing, generating a feature vector of each word based on a result of the analysis and the parsing, and calculating a split estimation score for each word using the feature vector, and
• estimating the split position based on the split estimation score,
• the split estimation score for said each word is calculated using an LSTM model adapted to take into account a context before the word, a context after the word, or a context before and after the word, and
• the feature vector includes a word index assigned to said each word from the beginning of a sentence, a character length, a word class, and a result of dependency parsing.

A non-transitory computer readable storage medium storing computer executable instructions which, upon being executed by a processor, control the processor to perform processes comprising:

• sequentially acquiring texts from voice recognition results;
• storing the texts as archive datasets;
• estimating a split position and a concatenation position of the texts from one or more of the archive datasets;
• generating a subtitle text from said one or more of the archive datasets based on the split position and the concatenation position;
• updating the archive datasets based on the split position and the concatenation position;
• and
• displaying the subtitle text on a display,
• wherein:
• the estimating comprises:
• continuing estimation of the concatenation position so as to concatenate the texts in the archive datasets, until a number of characters of an obtained concatenated text reaches a threshold character number or greater,
• subjecting the concatenated text to one or both of a morphological analysis and a dependency parsing, generating a feature vector of each word based on a result of the analysis and the parsing, and calculating a split estimation score for each word using the feature vector, and
• estimating the split position based on the split estimation score,
• the split estimation score for said each word is calculated using an LSTM model adapted to take into account a context before the word, a context after the word, or a context before and after the word, and
• the feature vector includes a word index assigned to said each word from the beginning of a sentence, a character length, a word class, and a result of dependency parsing.