Learning a verification model for speech recognition based on extracted recognition and language feature information

A speech processing method comprising:

• extracting recognition feature information having a characteristic of recognition result data obtained by performing a speech recognition process on an inputted speech, from said recognition result data by a speech processing apparatus, said recognition feature information being speech recognition result data for learning which includes plural recognition hypotheses;
• extracting language feature information having a characteristic of a pre-registered language resource from said pre-registered language resource by said speech processing apparatus, said pre-registered language resource including document data, sentences, text data, word sequences, or dictionaries, the extracted language feature information including linguistic characteristics included in an existing word sequence, or importance or similarity of a document;
• and
• obtaining a verification model by a learning process based on the extracted recognition feature information and language feature information by said speech processing apparatus, the obtained verification model being used to verify a speech recognition result data which is inputted as a verification target to a speech recognition system, wherein said obtained verification model is a discriminative model, said discriminative model being indicative of a correct and false label or degree of importance according to use.

A speech processing apparatus comprising:

• means for extracting recognition feature information having a characteristic of recognition result data obtained by performing a speech recognition process on an inputted speech, from said recognition result data, said recognition feature information being speech recognition result data for learning which includes plural recognition hypotheses;
• means for extracting language feature information having a characteristic of a pre-registered language resource from said pre-registered language resource, said pre-registered language resource including document data, sentences, text data, word sequences, or dictionaries, the extracted language feature information including linguistic characteristics included in an existing word sequence, or importance or similarity of a document;
• and
• means for obtaining a verification model by a learning process based on the extracted recognition feature information and language feature information, the obtained verification model being used to verify a speech recognition result data which is inputted as a verification target to a speech recognition system, wherein said obtained verification model is a discriminative model, said discriminative model being indicative of a correct and false label or degree of importance according to use.