Speech recognition apparatus and method thereof

A speech recognition method carried out by the speech recognition apparatus including a model storage unit having a plurality of phrases expressed on basis of grammar and one or more continuous phrase segments obtained by dividing the respective phrase, the model storage unit configured to store state transition models which express time series changes of the speech features for respective phrase segments as state-to-state transition relating to the speech features;

• a first grammar storage unit configured to store grammar segments relating to one or more continuous phrase segments which belong to each of the phrases;
• and
• a second grammar storage unit configured to store at all the steps or one step of grammar segments transferred from the first grammar storage unit and to be able to read out information stored therein in a reading time shorter than that required for the first grammar storage unit;
• the method comprising:
• a generating step of generating sequences of speech features from characteristics of entered speech for respective frames having an arbitrary temporal width;
• a first decoding step of obtaining forward probabilities of respective states of the state transition models for the sequence of speech features generated in the generating step with respect to each of the frames, by referring to the grammar segments stored in the second grammar storage unit and the state transition models stored in the model storage unit;
• a grammar transferring step of transferring a trailing grammar segment relating to a trailing phrase segment which trails one of said continuous phrase segments, from the first grammar storage unit to the second grammar storage unit when the forward probability of final state among said states of the state transition models is obtained in the first decoding step;
• a second decoding step of obtaining the forward probabilities of the respective states of the state transition models for a sequence of trailing speech features as the sequence of speech features for the trailing segment as generated by the generating unit with respect to each of the frames, continuously after the speech feature sequences, by referring to the grammar segments stored in the second grammar storage unit and the state transition models stored in the model storage unit;
• a third decoding step of obtaining the forward probabilities of the respective states of the state transition models for the trailing speech feature sequences for the respective frames, by referring to the trailing grammar segment transferred to the second grammar storage unit and the state transition models stored in the model storage unit;
• a recognition controlling step of (1) carrying out recognition for the respective phrases, (2) activating the first decoding step until the transfer of the trailing grammar segment is started, (3) activating the second decoding step in parallel to the transfer from the start to the completion of the transfer, (4) activating the third decoding step upon completion of the transfer, and (5) repeating the operations from (2) to (4) until all the operations for the phrase segments belonging to the respective phrases to obtain final forward probabilities for the respective phrases;
• and
• a recognizing step of outputting the phrase which give the highest forward probability from among the respective final forward probabilities of the plurality of phrases as a result of recognition of the speech feature sequence.

A speech recognition program stored in a non-transitory computer-readable medium and being executable in a computer,

• the computer comprising:
• a model storage unit having a plurality of phrases expressed on basis of grammar and one or more continuous phrase segments obtained by dividing the respective phrase segments, the model storage unit configured to store state transition models which express time series changes of the speech features for respective phrase segments as state-to-state transition relating to the speech features;
• a first grammar storage unit configured to store respective instructed grammar segments relating to one or more continuous phrase segments which belong to each of the phrases;
• and
• a second grammar storage unit configured to store at least part of the grammar segments transferred from the first grammar storage unit and to be able to read out information stored therein in a reading time shorter than that required for the first grammar storage unit;
• the program comprising:
• a generating instruction of generating sequences of speech features from characteristics of entered speech for respective frames having an arbitrary temporal width;
• a first decoding instruction of obtaining forward probabilities of respective states of the state transition models for the sequence of speech features generated by the generating instruction with respect to each of the frames, by referring to the grammar segments stored in the second grammar storage unit and the state transition models stored in the model storage unit;
• a grammar transferring instruction of transferring a trailing grammar segment relating to a trailing phrase segment which trails one of said continuous phrase segments, from the first grammar storage unit to the second grammar storage unit when the forward probability of final state among said states of the state transition models is obtained by the first decoding instruction;
• a second decoding instruction of obtaining the forward probabilities of the respective states of the state transition models for a sequence of trailing speech features as the sequence of speech features for the trailing segment as generated the generating instruction with respect to each of the frames, continuously after the speech feature sequences, by referring to the grammar segments stored in the second grammar storage unit and the state transition models stored in the model storage unit;
• a third decoding instruction of obtaining the forward probabilities of the respective states of the state transition models for the trailing speech feature sequences for the respective frames, by referring to the trailing grammar segment transferred to the second grammar storage unit and the state transition models stored in the model storage unit;
• a recognition controlling instruction of (1) carrying out recognition for the respective phrases, (2) activating the first decoding instruction until the transfer of the trailing grammar segment is started, (3) activating the second decoding instruction in parallel to the transfer from the start to the completion of the transfer, (4) activating the third decoding instruction upon completion of the transfer, and (5) repeating the operations from (2) to (4) until all the operations for the phrase segments belonging to the respective phrases to obtain final forward probabilities for the respective phrases;
• and
• a recognizing instruction of outputting the phrase which gives the highest forward probability from among the respective final forward probabilities of the plurality of phrases as a result of recognition of the speech feature sequence.