Multi-stage speech recognition apparatus and method

1. A multi-stage speech recognition apparatus comprising:

at least one processor that executes:

a first speech recognition unit performing initial speech recognition on a feature vector, which is extracted from an input speech signal, and generating a plurality of candidate words; and

a second speech recognition unit rescoring the candidate words, which are provided by the first speech recognition unit, using a temporal posterior feature vector reflected from time-varying voice characteristics and extracted from the feature vector, and outputting a word having a highest score as a final recognition result;

wherein the temporal posterior feature vector is a split-temporal context (STC)-TRAP feature vector comprising a left, center, and right context part.

2. The apparatus of claim 1 , wherein the first speech recognition unit comprises:

a first feature extractor extracting a spectrum feature vector from the speech signal; and

a recognizer performing the initial speech recognition using the spectrum feature vector.

3. The apparatus of claim 1 , wherein the second speech recognition unit comprises:

a second feature extractor extracting the temporal posterior feature vector from the feature vector; and

a rescorer performing forced alignment of the candidate words, which are provided by the first speech recognition unit, using the temporal posterior feature vector.

4. The apparatus of claim 1 , wherein the split-temporal context (STC)-TRAP feature vector is obtained by inputting feature vectors of adjacent frames placed before and after a current frame, which is to be converted, to a left context neural network, a center context neural network, and a right context neural network, for respective bands, and integrating outputs of the left context neural network, the center context neural network, and the right context neural network.

5. A multi-stage speech recognition method comprising:

performing, by at least one processor, initial speech recognition on a feature vector, which is extracted from an input speech signal, and generating a plurality of candidate words; and

rescoring, by the at least one processor, the candidate words, which are obtained from the initial speech recognition, using a temporal posterior feature vector reflected from time- varying voice characteristics and extracted from the speech signal, and outputting a word having a highest score as a final recognition result;

wherein the temporal posterior feature vector is a split-temporal context (STC)-TRAP feature vector comprising a left, center, and right context part.

6. The method of claim 5 , wherein the performing of the initial speech recognition comprises:

extracting a spectrum feature vector from the speech signal; and

performing the initial speech recognition using the spectrum feature vector.

7. The method of claim 5 , wherein the rescoring of the candidate words comprises:

extracting the temporal posterior feature vector from the feature vector; and

performing forced alignment of the candidate words, which are obtained from the initial speech recognition, using the temporal posterior feature vector.

8. The method of claim 5 , wherein the split-temporal context (STC)-TRAP feature vector is obtained by inputting feature vectors of adjacent frames placed before and after a current frame, which is to be converted, to a left context neural network, a center context neural network, and a right context neural network, for respective bands, and integrating outputs of the left context neural network, the center context neural network, and the right context neural network.

9. A non-transitory computer-readable recording medium storing a program to control at least one processing element to implement the method of claim 5 .