Enhancing hybrid self-attention structure with relative-position-aware bias for speech synthesis

1. A method of performing speech synthesis, the method comprising:

encoding character embeddings, using any one or any combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs);

applying a relative-position-aware self attention function to each of the character embeddings and an input mel-scale spectrogram;

encoding the character embeddings to which the relative-position-aware self attention function is applied, by applying a layer normalization, a sequence of feed forward transformations, and a learned linear transformation to the character embeddings to which the relative-position-aware self attention function is applied;

concatenating the encoded character embeddings and the encoded character embeddings to which the relative-position-aware self attention function is applied, to generate an encoder output;

applying a multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied; and

predicting an output mel-scale spectrogram, based on the encoder output and the input mel-scale spectrogram to which the multi-head attention function is applied.

2. The method of claim 1 , further comprising generating a speech waveform, based on the predicted output mel-scale spectrogram.

3. The method of claim 1 , wherein the applying the multi-head attention function comprises:

modifying the multi-head attention function to enhance local relations, using an edge representation for a matrix; and

applying the modified multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied.

4. The method of claim 1 , wherein the applying the multi-head attention function comprises:

modifying the multi-head attention function to enhance local relations, by clipping, in both directions, a maximum relative position of an edge representation for a matrix; and

applying the modified multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied.

5. The method of claim 1 , further comprising applying the layer normalization, the sequence of feed forward transformations, and the learned linear transformation to the encoder output and the input mel-scale spectrogram to which the multi-head attention function is applied, to generate a decoder output,

wherein the predicting the output mel-scale spectrogram comprises predicting the output mel-scale spectrogram, based on the decoder output.

6. The method of claim 1 , wherein the character embeddings are related to one or more linguistic features of input text, and

the input mel-scale spectrogram and the output mel-scale spectrogram are related to one or more acoustic features of the one or more linguistic features.

7. An apparatus for performing speech synthesis, the apparatus comprising:

at least one memory configured to store program code; and

at least one processor configured to read the program code and operate as instructed by the program code, the program code including:

first encoding code configured to cause the at least one processor to encode character embeddings, using any one or any combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs);

first applying code configured to cause the at least one processor to apply a relative-position-aware self attention function to each of the character embeddings and an input mel-scale spectrogram;

second encoding code configured to cause the at least one processor to encode the character embeddings to which the relative-position-aware self attention function is applied, by applying a layer normalization, a sequence of feed forward transformations, and a learned linear transformation to the character embeddings to which the relative-position-aware self attention function is applied;

concatenating code configured to cause the at least one processor to concatenate the encoded character embeddings and the encoded character embeddings to which the relative-position-aware self attention function is applied, to generate an encoder output;

second applying code configured to cause the at least one processor to apply a multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied; and

predicting code configured to cause the at least one processor to predict an output mel-scale spectrogram, based on the encoder output and the input mel-scale spectrogram to which the multi-head attention function is applied.

8. The apparatus of claim 7 , further comprising generating code configured to cause the at least one processor to generate a speech waveform, based on the predicted output mel-scale spectrogram.

9. The apparatus of claim 7 , wherein the second applying code is further configured to cause the at least one processor to:

modify the multi-head attention function to enhance local relations, using an edge representation for a matrix; and

apply the modified multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied.

10. The apparatus of claim 7 , wherein the second applying code is further configured to cause the at least one processor to:

modify the multi-head attention function to enhance local relations, by clipping, in both directions, a maximum relative position of an edge representation for a matrix; and

apply the modified multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied.

11. The apparatus of claim 7 , further comprising decoding code configured to cause the at least one processor to apply the layer normalization, the sequence of feed forward transformations, and the learned linear transformation to the encoder output and the input mel-scale spectrogram to which the multi-head attention function is applied, to generate a decoder output,

wherein the predicting code is further configured to cause the at least one processor to predict the output mel-scale spectrogram, based on the decoder output.

12. The apparatus of claim 7 , wherein the character embeddings are related to one or more linguistic features of input text, and

the input mel-scale spectrogram and the output mel-scale spectrogram are related to one or more acoustic features of the one or more linguistic features.

13. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor of a device, cause the at least one processor to:

encode character embeddings, using any one or any combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs);

apply a relative-position-aware self attention function to each of the character embeddings and an input mel-scale spectrogram;

encode the character embeddings to which the relative-position-aware self attention function is applied, by applying a layer normalization, a sequence of feed forward transformations, and a learned linear transformation to the character embeddings to which the relative-position-aware self attention function is applied;

concatenate the encoded character embeddings and the encoded character embeddings to which the relative-position-aware self attention function is applied, to generate an encoder output;

apply a multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied; and

predict an output mel-scale spectrogram, based on the encoder output and the input mel-scale spectrogram to which the multi-head attention function is applied.

14. The non-transitory computer-readable medium of claim 13 , wherein the instructions further cause the at least one processor to generate a speech waveform, based on the predicted output mel-scale spectrogram.

15. The non-transitory computer-readable medium of claim 13 , wherein the instructions further cause the at least one processor to:

modify the multi-head attention function to enhance local relations, using an edge representation for a matrix; and

apply the modified multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied.

16. The non-transitory computer-readable medium of claim 13 , wherein the instructions further cause the at least one processor to:

modify the multi-head attention function to enhance local relations, by clipping, in both directions, a maximum relative position of an edge representation for a matrix; and

apply the modified multi-head attention function to the encoder output and the input mel-scale spectrogram to which the relative-position-aware self attention function is applied.

17. The non-transitory computer-readable medium of claim 13 , wherein the instructions further cause the at least one processor to:

apply the layer normalization, the sequence of feed forward transformations, and the learned linear transformation to the encoder output and the input mel-scale spectrogram to which the multi-head attention function is applied, to generate a decoder output; and

predict the output mel-scale spectrogram, based on the decoder output.