Attention-based clockwork hierarchical variational encoder
A method for representing an intended prosody in synthesized speech includes receiving a text utterance having at least one word, and selecting an utterance embedding for the text utterance. Each word in the text utterance has at least one syllable and each syllable has at least one phoneme. The utterance embedding represents an intended prosody. For each syllable, using the selected utterance embedding, the method also includes: predicting a duration of the syllable by decoding a prosodic syllable embedding for the syllable based on attention by an attention mechanism to linguistic features of each phoneme of the syllable and generating a plurality of fixed-length predicted frames based on the predicted duration for the syllable.
1. A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:
receiving a text utterance having at least one word, each word having at least one syllable, each syllable having at least one phoneme;
selecting an utterance embedding for the text utterance, the utterance embedding representing an intended prosody; and
for each corresponding syllable, using the selected utterance embedding:
predicting a duration of the corresponding syllable by decoding a prosodic syllable embedding for the corresponding syllable based on attention by an attention mechanism to linguistic features of each phoneme of an adjacent syllable to the corresponding syllable; and
generating a plurality of fixed-length predicted frames based on the predicted duration for the corresponding syllable.
2. The method of claim 1 , wherein the operations further comprise:
predicting a pitch contour of the corresponding syllable based on the predicted duration of the corresponding syllable, and
wherein the plurality of fixed-length predicted frames comprise fixed-length predicted pitch frames, each fixed-length predicted pitch frame representing part of the predicted pitch contour of the corresponding syllable.
3. The method of claim 1 , wherein the operations further comprise, for each corresponding syllable, using the selected utterance embedding:
predicting an energy contour of the corresponding syllable based on the predicted duration of the corresponding syllable; and
generating a plurality of fixed-length predicted energy frames based on the predicted duration of the corresponding syllable, each fixed-length energy frame representing the predicted energy contour of the corresponding syllable.
4. The method of claim 1 , wherein the plurality of fixed length predicted frames comprise fixed-length predicted spectral frames for the corresponding syllable.
5. The method of claim 1 , wherein a network representing a hierarchical linguistic structure of the text utterance comprises:
a first level including each word of the text utterance;
a second level including each syllable of the text utterance; and
a third level including each fixed-length predicted frame for each syllable of the text utterance.
6. The method of claim 1 , wherein predicting the duration of the corresponding syllable comprises:
for each phoneme associated with the adjacent syllable to the corresponding syllable:
encoding one or more linguistic features of a corresponding phoneme;
inputting the encoded one or more linguistic features into the attention mechanism; and
applying the attention of the attention mechanism to the prosodic syllable embedding.
7. The method of claim 1 , wherein the adjacent syllable to the corresponding syllable comprises one of a preceding adjacent syllable that precedes the corresponding syllable or a subsequent adjacent syllable that is subsequent to the corresponding syllable.
8. The method of claim 7 , wherein predicting the duration of the corresponding syllable by decoding the prosodic syllable embedding for the corresponding syllable is further based on attention by the attention mechanism to linguistic features of each phoneme of another adjacent syllable to the corresponding syllable, the another adjacent syllable to the corresponding syllable comprising the other one of the preceding adjacent syllable or the subsequent adjacent syllable.
9. The method of claim 1 , wherein the operations further comprise:
receiving training data including a plurality of reference audio signals, each reference audio signal comprising a spoken utterance of human speech and having a corresponding prosody; and
training a deep neural network for a prosody model by encoding each reference audio signal into a corresponding fixed-length utterance embedding representing the corresponding prosody of the reference audio signal.
10. The method of claim 9 , wherein the operations further comprise generating the selected utterance embedding by encoding linguistic features for a plurality of linguistic units with a frame-based syllable embedding and a phone feature-based syllable embedding.
11. The method of claim 1 , wherein the utterance embedding comprises a fixed-length numerical vector.
12. The method of claim 1 , wherein the attention of the attention mechanism comprises location-based attention.
13. The method of claim 12 , wherein the location-based attention comprises monotonically shifting, location sensitive attention, the monotonically shifting, location sensitive attention defined by a window of phoneme information for a respective syllable.
14. The method of claim 1 , wherein the attention mechanism comprises a transformer.
15. A system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware, the memory hardware storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:
receiving a text utterance having at least one word, each word having at least one syllable, each syllable having at least one phoneme;
selecting an utterance embedding for the text utterance, the utterance embedding representing an intended prosody; and
for each corresponding syllable, using the selected utterance embedding:
predicting a duration of the corresponding syllable by decoding a prosodic syllable embedding for the corresponding syllable based on attention by an attention mechanism to linguistic features of each phoneme of an adjacent syllable to the corresponding syllable; and
generating a plurality of fixed-length predicted frames based on the predicted duration for the corresponding syllable.
16. The system of claim 15 , wherein the operations further comprise:
predicting a pitch contour of the corresponding syllable based on the predicted duration of the corresponding syllable, and
wherein the plurality of fixed-length predicted frames comprise fixed-length predicted pitch frames, each fixed-length predicted pitch frame representing part of the predicted pitch contour of the corresponding syllable.
17. The system of claim 15 , wherein the operations further comprise, for each corresponding syllable, using the selected utterance embedding:
predicting an energy contour of the corresponding syllable based on the predicted duration of the corresponding syllable; and
generating a plurality of fixed-length predicted energy frames based on the predicted duration of the corresponding syllable, each fixed-length energy frame representing the predicted energy contour of the corresponding syllable.
18. The system of claim 15 , wherein the plurality of fixed length predicted frames comprise fixed-length predicted spectral frames for the corresponding syllable.
19. The system of claim 15 , wherein a network representing a hierarchical linguistic structure of the text utterance comprises:
a first level including each word of the text utterance;
a second level including each syllable of the text utterance; and
a third level including each fixed-length predicted frame for each syllable of the text utterance.
20. The system of claim 15 , wherein predicting the duration of the corresponding syllable comprises:
for each phoneme associated with the adjacent syllable to the corresponding syllable:
encoding one or more linguistic features of a corresponding phoneme;
inputting the encoded one or more linguistic features into the attention mechanism; and
applying the attention of the attention mechanism to the prosodic syllable embedding.
21. The system of claim 15 , wherein the adjacent syllable to the corresponding syllable comprises one of a preceding adjacent syllable that precedes the corresponding syllable or a subsequent adjacent syllable that is subsequent to the corresponding syllable.
22. The system of claim 21 , wherein predicting the duration of the corresponding syllable by decoding the prosodic syllable embedding for the corresponding syllable is further based on attention by the attention mechanism to linguistic features of each phoneme of another adjacent syllable to the corresponding syllable, the another adjacent syllable to the corresponding syllable comprising the other one of the preceding adjacent syllable or the subsequent adjacent syllable.
23. The system of claim 15 , wherein the operations further comprise:
receiving training data including a plurality of reference audio signals, each reference audio signal comprising a spoken utterance of human speech and having a corresponding prosody; and
training a deep neural network for a prosody model by encoding each reference audio signal into a corresponding fixed-length utterance embedding representing the corresponding prosody of the reference audio signal.
24. The system of claim 23 , wherein the operations further comprise generating the selected utterance embedding by encoding linguistic features for a plurality of linguistic units with a frame-based syllable embedding and a phone feature-based syllable embedding.
25. The system of claim 15 , wherein the utterance embedding comprises a fixed-length numerical vector.
26. The system of claim 15 , wherein the attention of the attention mechanism comprises location-based attention.
27. The system of claim 26 , wherein the location-based attention comprises monotonically shifting, location sensitive attention, the monotonically shifting, location sensitive attention defined by a window of phoneme information for a respective syllable.
28. The system of claim 15 , wherein the attention mechanism comprises a transformer.