Speech recognition system and method

1. A speech recognition method which allows phones to be automatically generated, comprising:

unsupervisedly learning a feature vector of speech data;

generating a phone set by clustering acoustic features selected based on an unsupervised learning result;

allocating a sequence of phones to the speech data on the basis of the generated phone set;

generating an acoustic model on the basis of the sequence of phones and the speech data to which the sequence of phones is allocated; and

generating a speech recognition result by a speech recognition decoder using the acoustic model and a language network.

2. The speech recognition method of claim 1 , wherein the speech data is untranscribed speech data.

3. The speech recognition method of claim 2 , wherein the unsupervisedly learning of the feature vector of the speech data includes:

extracting the feature vector from the speech data;

unsupervisedly learning the extracted feature vector; and

generating an artificial neural network including an acoustic pattern that corresponds to the feature vector based on the unsupervised learning result.

4. The speech recognition method of claim 3 , wherein the extracting of the feature vector from the speech data includes:

converting the speech data into a spectrogram;

generating a first feature vector by converting the speech data, which is converted into the spectrogram, into the first feature vector using a Mel-scale filter bank in a predetermined time frame unit; and

splicing left and right windows of the first feature vector according to a predetermined number of frames and generating a second feature vector,

wherein the generated second feature vector is extracted as the feature vector.

5. The speech recognition method of claim 4 , wherein the unsupervisedly learning of the extracted feature vector includes unsupervisedly learning the feature vector by arranging the extracted feature vector on an input node and an output node of a stacked autoencoder.

6. The speech recognition method of claim 3 , wherein the extracting of the feature vector from the speech data includes:

converting the speech data into a spectrogram; and

generating a feature matrix by grouping the speech data converted into the spectrogram into 2-dimensional units of x frames,

wherein the generated feature matrix is extracted as the feature vector.

7. The speech recognition method of claim 6 , wherein the unsupervisedly learning of the extracted feature vector includes unsupervisedly learning the feature vector by arranging the extracted feature vector on an input node and an output node of a convolutional autoencoder.

8. The speech recognition method of claim 3 , wherein the generating of the phone set by clustering the acoustic features selected based on the unsupervised learning result includes generating the phone set by listing an output value for each piece of input data of the artificial neural network.

9. The speech recognition method of claim 8 , wherein the generating of the phone set includes:

representing the output value for each piece of input data as a vector and listing vectors;

extracting vectors whose distance therebetween is equal to or smaller than a specific bound value from the listed vectors on the basis of vector clustering;

generating a group vector by averaging the extracted vectors; and

generating the phone set on the basis of the listed vectors and the generated group vector.

10. The speech recognition method of claim 8 , wherein the generating of the phone set includes:

listing an index of a node using the output value for each piece of input data; and

generating the phone set by performing the clustering, focusing on indices whose frequencies of being output are equal to or greater than a predetermined number of times among the listed indices.

11. The speech recognition method of claim 8 , wherein the allocating of the sequence of phones to the speech data includes:

listing a candidate sequence of phones on the basis of the artificial neural network; and

allocating a final sequence of phones to the speech data by extracting the final sequence of phones on the basis of the generated phone set and the candidate sequence of phones.

12. The speech recognition method of claim 11 , wherein the generating of the acoustic model includes:

generating a context-independent phone sequence model using the sequence of phones and the speech data to which the sequence of phones is re-allocated;

generating a context-dependent tree on the basis of the context-independent phone sequence model and a combination according to the context of the sequence of phones;

defining a context-dependent phone state on the basis of the context-dependent tree;

allocating the defined context-dependent phone state to the speech data using the sequence of phones; and

training context-dependent phone state models on the basis of information on the allocated context-dependent phone state and the speech data.

13. The speech recognition method of claim 12 , wherein the training of the context-dependent phone state models includes:

re-allocating the sequence of phones to the speech data using the trained context-dependent phone state models; and

re-training the context-dependent phone state models on the basis of the speech data and information on context-dependent phone state derived from the re-allocated sequence of phones.

14. The speech recognition method of claim 1 , further comprising generating a space-separated word segment-based pronunciation dictionary on the basis of transcribed speech data,

wherein the generating of the space-separated word segment-based pronunciation dictionary includes:

splitting the transcribed speech data into sections in units of word segment;

allocating the sequence of phones to the transcribed speech data which is split into the word segments;

arranging the sequence of phones to a corresponding word segment of the transcribed speech data to which the sequence of phones is allocated;

refining the arranged sequence of phones on the basis of a time axis and the number of frames; and

generating the space-separated word segment-based pronunciation dictionary on the basis of the refined sequence of phones.

15. The speech recognition method of claim 14 , wherein the generating of the space-separated word segment-based pronunciation dictionary includes generating the pronunciation dictionary in units of syllables or partial word segments generated by splitting a word segment of the transcribed speech data.

16. The speech recognition method of claim 15 , further comprising:

linking a word segment-based language model generated by a learning corpus with the generated pronunciation dictionary; and

generating a language network according to the linking result.

17. The speech recognition method of claim 16 , wherein the linking of the word segment-based language model with the pronunciation dictionary includes expanding the word segment-based pronunciation dictionary by linking words of the language model which are not included in the word segment-based pronunciation dictionary with the pronunciation dictionary in units of partial word segments or syllables.

18. The speech recognition method of claim 16 , further comprising:

extracting a feature vector from speech data input by a user;

inputting the feature vector to the speech recognition decoder to which the generated acoustic model and the language network are applied; and

generating the speech recognition result by extracting a sequence of words of the input speech data on the basis of a result of inputting.

19. A speech recognition system which allows phones to be generated, comprising:

a memory configured to store a program for speech recognition; and

a processor configured to execute the program stored in the memory,

wherein the processor performs unsupervised learning by extracting a feature vector from untranscribed speech data upon executing the program, generates a phone set by clustering acoustic features selected based on an unsupervised learning result, allocates a sequence of phones to the speech data on the basis of the generated phone set, generates an acoustic model on the basis of the sequence of phones and the speech data to which the sequence of phones is allocated, and generates a speech recognition result using the acoustic model and a language network.