Method for generating learning data for speech recognition error detection

1. A method for generating data, the method performed by a computing device, the method comprising:

segmenting text data generated based on speech information into a token unit;

generating a first feature vector based on the text data segmented into the token unit, and generating a first label vector corresponding to the generated first feature vector;

generating a second feature vector and a second label vector by performing mix-up for each of the generated first feature vector and the generated first label vector;

determining a prediction label by inputting the generated second feature vector into a neural network model, and

training the neural network model based on the determined prediction label and the second label vector.

2. The method of claim 1 , wherein the data generation is to train a neural network model for speech recognition error detection, and

wherein the mix-up includes manifold-mix-up.

3. The method of claim 1 , wherein the generating of the second feature vector and the second label vector by performing the mix-up for each of the generated first feature vector and the generated first label vector includes:

performing a shuffle by applying a same index to the first feature vector and the first label vector.

4. The method of claim 3 , wherein the performing of the shuffle by applying the same index to the first feature vector and the first label vector includes:

performing the shuffle based on random permutation for the first feature vector, and performing the shuffle for the first label vector based on a same random permutation.

5. The method of claim 3 , wherein the generating of the second feature vector and the second label vector by performing the mix-up for each of the generated first feature vector and the generated first label vector further includes:

generating the second feature vector by performing the mix-up between the first feature vector and a shuffled version of the first feature vector; and

generating the second label vector by performing the mix-up between the first label vector and the shuffled version of the first label vector.

6. The method of claim 1 , wherein the determining of the prediction label by inputting the generated second feature vector into the neural network model includes:

determining a label by predicting whether an error occurs based on whether an output value of the neural network model exceeding a set threshold value.

7. The method of claim 6 , wherein the training of the neural network model based on the determined prediction label and the second label vector includes:

setting the second label vector to a ground truth, and training the neural network model by making the prediction label correspond to the second label vector.

8. The method of claim 7 , wherein the setting of the second label vector to the ground truth, and training of the neural network model by making the prediction label correspond to the second label vector includes:

calculating an error between the prediction label and the second label vector by using a loss function and training the neural network model based on the calculated error.

9. A computing device comprising:

at least one processor; and

a memory,

wherein the at least one processor is configured to:

segment text data generated based on speech information into a token unit;

generate a first feature vector based on the text data segmented into the token unit, and generate a first label vector corresponding to the generated first feature vector,

generate a second feature vector and a second label vector by performing mix-up for each of the generated first feature vector and the generated first label vector,

determine a prediction label by inputting the generated second feature vector into a neural network model, and

train the neural network model based on the determined prediction label and the second label vector.

10. A method for detecting a speech recognition error, the method performed by a computing device, the method comprising:

segmenting text data generated based on speech information into a token unit; and

performing speech recognition error detection based on the segmented tokens by using a pre-learned neural network model,

wherein the pre-learned neural network model corresponds to a model learned based on data generation,

wherein the data generation includes generating a second feature vector and a second label vector by performing mix-up for each of a first feature vector generated based on a plurality of tokens and a first label vector corresponding to the first feature vector, and

wherein the pre-learned neural network model corresponds to a model learned based on a predicted label generated based on the second feature vector and the second label vector.