Recognition result output device, recognition result output method, and computer program product

1. A recognition result output device comprising:

first storage configured to store a language model for speech recognition;

second storage configured to store an acoustic model for speech recognition; and

processing circuitry coupled to the first storage and the second storage, and configured to function as a phonetic sequence acquirer, a first feature converter, a first grapheme converter, an input speech acquirer, a second feature converter, and a second graphene converter, wherein

the processing circuitry determines whether input is a phonetic sequence or a speech;

when the input is the phonetic sequence, the phonetic sequence acquirer acquires the phonetic sequence,

the first feature converter converts the phonetic sequence into a phonetic sequence feature vector, the phonetic sequence feature vector including a plurality of acoustic scores, each of the plurality of acoustic scores being an acoustic score of a phoneme included in the phonetic sequence,

the first grapheme converter converts the phonetic sequence feature vector into graphemes using the language model, and outputs the graphemes,

when the input is the speech, the input speech acquirer acquires the speech,

the second feature converter converts a speech waveform of the acquired speech into a speech feature vector for speech recognition,

the second grapheme converter converts the speech feature vector into graphemes using the language model and the acoustic model, and

the recognition result output device further comprises a display configured to display the output graphemes.

2. The device according to claim 1 , wherein the phonetic sequence feature vector is an acoustic score vector.

3. The device according to claim 1 , wherein the phonetic sequence feature vector is a phoneme state acoustic score vector in which an element of a phoneme state acoustic score corresponding to a phonetic sequence in a phoneme state acoustic score vector sequence is set to be higher than other phoneme state acoustic scores.

4. The device according to claim 1 , wherein

the acoustic model is a Gaussian distribution acoustic model, and

the phonetic sequence feature vector has average values of a plurality of dimensions of a mixture Gaussian acoustic model that represents a phonetic sequence state, as elements.

5. A recognition result output method employed in a recognition result output device comprising:

determining whether input is a phonetic sequence or a speech;

acquiring, when the input is the phonetic sequence, the phonetic sequence;

converting the acquired phonetic sequence into a phonetic sequence feature vector, the phonetic sequence feature vector including a plurality of acoustic scores, each of the plurality of acoustic scores being an acoustic score of a phoneme included in the phonetic sequence;

converting the phonetic sequence feature vector into graphemes using a language model that has language statistical information for speech recognition;

outputting the graphemes,

acquiring, when the input is the speech, the speech;

converting a speech waveform of the acquired speech into a speech feature vector for speech recognition;

converting the speech feature vector into graphemes using the language model and an acoustic model, and

displaying the output graphemes.

6. The method according to claim 5 , wherein the phonetic sequence feature vector is an acoustic score vector.

7. The method according to claim 5 , wherein the phonetic sequence feature vector is a phoneme state acoustic score vector in which an element of a phoneme state acoustic score corresponding to a phonetic sequence in a phoneme state acoustic score vector sequence is set to be higher than other phoneme state acoustic scores.

8. The method according to claim 5 , wherein

the acoustic model is a Gaussian distribution acoustic model, and

the phonetic sequence feature vector has average values of a plurality of dimensions of a mixture Gaussian acoustic model that represents a phonetic sequence state, as elements.

9. A computer program product comprising a non-transitory computer-readable medium containing a program executed by a computer, the program causing the computer to execute:

determining whether input is a phonetic sequence or a speech;

acquiring, when the input is the phonetic sequence, the phonetic sequence;

converting the acquired phonetic sequence into a phonetic sequence feature vector, the phonetic sequence feature vector including a plurality of acoustic scores, each of the plurality of acoustic scores being an acoustic score of a phoneme included in the phonetic sequence;

converting the phonetic sequence feature vector into graphemes using a language model that has language statistical information for speech recognition; and

outputting the graphemes,

acquiring, when the input is the speech, the speech;

converting a speech waveform of the acquired speech into a speech feature vector for speech recognition;

converting the speech feature vector into graphemes using the language model and an acoustic model, and

displaying the output graphemes.

10. The product according to claim 9 , wherein the phonetic sequence feature vector is an acoustic score vector.

11. The product according to claim 9 , wherein the phonetic sequence feature vector is a phoneme state acoustic score vector in which an element of a phoneme state acoustic score corresponding to a phonetic sequence in a phoneme state acoustic score vector sequence is set to be higher than other phoneme state acoustic scores.

12. The product according to claim 9 , wherein

the acoustic model is a Gaussian distribution acoustic model, and

the phonetic sequence feature vector has average values of a plurality of dimensions of a mixture Gaussian acoustic model that represents a phonetic sequence state, as elements.