Synthetic speech processing

1. A computer-implemented method for generating synthesized speech, the method comprising:

receiving text data representing content to be transformed into synthetic speech;

processing, using a sequence-to-sequence model, the text data to determine Mel-spectrogram data representing a characteristic of the synthetic speech;

processing the Mel-spectrogram data to determine amplitude data corresponding to the synthetic speech;

determining, using an affine coupling layer of a normalizing flow decoder and the amplitude data, a network weight of the normalizing flow decoder;

processing, using the normalizing flow decoder and the network weight, at least a portion of the Mel-spectrogram data to determine phase data representing the characteristic;

processing, using an inverse Fourier transform component, the Mel-spectrogram data and the phase data to determine audio data representing the synthetic speech; and

causing output of audio corresponding to the audio data.

2. The computer-implemented method of claim 1 , further comprising:

determining second text data representing second speech;

determining second audio data representing the second speech; and

processing, using a normalizing flow encoder, the second text data and the second audio data to determine a Gaussian distribution,

wherein the phase data is based at least in part on the Gaussian distribution.

3. A computer-implemented method comprising:

receiving first data representing content to be synthesized as audio data;

processing the first data to determine second data representing a power value of the audio data;

processing, using a decoder, at least a portion of the second data to determine third data representing a phase value of the audio data; and

processing, using a first component, the second data and the third data to determine the audio data representing the content as synthesized speech.

4. The computer-implemented method of claim 3 , further comprising:

processing the second data to determine amplitude data corresponding to the first data; and

determining, using an affine coupling layer of the decoder and the amplitude data, a network weight of the decoder.

5. The computer-implemented method of claim 3 , further comprising:

determining second audio data representing an utterance; and

processing, using an encoder, the second audio data to determine a data distribution,

wherein the third data is based at least in part on the data distribution.

6. The computer-implemented method of claim 3 , further comprising at least one of:

processing the second data to determine amplitude data corresponding to the first data; and

determining a data distribution corresponding to the second data,

wherein the third data is based at least in part on the data distribution.

7. The computer-implemented method of claim 3 , further comprising:

determining fourth data representing a second power value of second audio data;

determining fifth data representing a second phase value of the second audio data;

processing, using a sequence-to-sequence model, the fourth data to determine a first data distribution; and

processing, using an encoder, the fifth data to determine a second data distribution.

8. The computer-implemented method of claim 3 , further comprising:

processing second text data to determine fourth data representing a second power value of second audio data;

processing, using an encoder, the fourth data to determine embedding data;

determining that a variance of a value of the embedding data satisfies a condition; and

processing, using the decoder, the value and at least a portion of the fourth data to determine a second phase value.

9. The computer-implemented method of claim 3 , further comprising:

processing, using an encoder, a first frame of power data to determine first embedding data;

processing, using the encoder, a second frame of the power data to determine second embedding data; and

processing, using a sequence-to-sequence model, the second embedding data to determine second audio data.

10. The computer-implemented method of claim 3 , further comprising:

receiving second data representing second content;

processing, using an encoder of a sequence-to-sequence model, the second data to determine embedding data; and

processing, using a second decoder, the embedding data to determine second audio data.

11. The computer-implemented method of claim 3 , further comprising:

receiving second audio data representing an utterance;

processing, using a feature extractor, the second audio data to determine a second power value of second audio data;

processing, using the decoder, the second power value to determine a second phase value of the second audio data; and

processing, using the first component, the second power value and the second phase value to determine third audio data that includes a representation of the utterance.

12. A system comprising:

at least one processor; and

at least one memory including instructions that, when executed by the at least one processor, cause the system to:

receive first data representing content to be synthesized as audio data;

process the first data to determine second data representing a power value of audio data;

process, using a decoder, at least a portion of the second data to determine third data representing a phase value of the audio data; and

process, using a first component, the second data and the third data to determine the audio data representing the content as synthesized speech.

13. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

process the second data to determine amplitude data corresponding to the first data; and

determine, using an affine coupling layer of the decoder and the amplitude data, a network weight of the decoder.

14. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

determine second audio data representing an utterance; and

process, using an flow encoder, the second audio data to determine a data distribution,

wherein the third data is based at least in part on the data distribution.

15. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

process the second data to determine amplitude data corresponding to the first data; and

determine a data distribution corresponding to the second data,

wherein the third data is based at least in part on the data distribution.

16. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

determine fourth data representing a second power value of second audio data;

determine fifth data representing a second phase value of the second audio data;

process, using a sequence-to-sequence model, the fourth data to determine a first data distribution; and

process, using an encoder, the fifth data to determine a second data distribution.

17. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

process second text data to determine fourth data representing a second power value of second audio data;

process, using an encoder, the fourth data to determine embedding data;

determine that a variance of a value of the embedding data satisfies a condition; and

process, using the decoder, the value and at least a portion of the fourth data to determine a second phase value.

18. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

process, using an encoder, a first frame of power data to determine first embedding data;

process, using the encoder, a second frame of the power data to determine second embedding data; and

process, using a sequence-to-sequence model, the second embedding data to determine second audio data.

19. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

receive second text data representing second content;

process, using an encoder of a sequence-to-sequence model, the second text data to determine embedding data; and

process, using a second decoder, the embedding data to determine second audio data.

20. The system of claim 12 , wherein the at least one memory further includes instructions that, when executed by the at least one processor, further cause the system to:

receive second audio data representing an utterance;

process, using a feature extractor, the second audio data to determine a second power value of second audio data;

process, using the decoder, the second power value to determine a second phase value; and

process, using the first component, the second power value and the second phase value to determine third audio data that includes a representation of the utterance.