Phonemes and graphemes for neural text-to-speech

1. A computer-implemented method when executed by data processing hardware causes the data processing hardware to perform operations comprising:

receiving, at an encoder of a speech synthesis model, a text input comprising a sequence of words represented as an input encoder embedding, the input encoder embedding comprising a plurality of tokens, the plurality of tokens comprising a first set of grapheme tokens representing the text input as respective graphemes and a second set of phoneme tokens representing the text input as respective phonemes, each grapheme token of the first set of grapheme tokens comprising a respective wordpiece sub-word of a respective word in the sequence of words, wherein each corresponding token of the plurality of tokens of the input encoder embedding represents a combination of:

a respective word position embedding for each respective word in the sequence of words, the respective word position embedding representing sub-word level positions for both one or more of the grapheme tokens from the first set of grapheme tokens that correspond to the respective word and one or more of the phoneme tokens from the second set of phoneme tokens that correspond to the respective word; and

a position embedding representing an overall index of position for each token of the plurality of tokens of the input encoder embedding;

for each respective phoneme token of the second set of phoneme tokens:

identifying, by the encoder, a respective word of the sequence of words corresponding to the respective phoneme token based on the respective word position embedding that represents the sub-word level position for the respective phoneme token that corresponds to the respective word; and

determining, by the encoder, a respective grapheme token representing the respective word of the sequence of words corresponding to the respective phoneme token by determining that the sub-word level position for the respective grapheme token that corresponds to the respective word is represented by the same respective word position embedding as the respective word position embedding representing the sub-word level position for the respective phoneme token; and

generating, by the encoder, an output encoder embedding based on a relationship between each respective phoneme token and the respective grapheme token determined to represent a same respective word as the respective phoneme token.

2. The method of claim 1 , wherein the combination representing each token of the plurality of tokens of the input encoder embedding further comprises:

one of a grapheme token embedding or a phoneme token embedding; and

a segment embedding.

3. The method of claim 1 , wherein the speech synthesis model comprises an attention mechanism in communication with the encoder.

4. The method of claim 1 , wherein the speech synthesis model comprises a duration-based upsampler in communication with the encoder.

5. The method of claim 1 , wherein the plurality of tokens of the input encoder embedding comprises a special token identifying a language of the input text.

6. The method of claim 1 , wherein the operations further comprise:

pre-training the encoder of the speech synthesis model by:

feeding the encoder a plurality of training examples, each training example represented as a sequence of training grapheme tokens corresponding to a training sequence of words and a sequence of training phoneme tokens corresponding to the same training sequence of words;

masking a training phoneme token from the sequence of training phoneme tokens for a respective word from the training sequence of words; and

masking a training grapheme token from the sequence of training phoneme tokens for the respective word from the training sequence of words.

7. The method of claim 1 , wherein:

the speech synthesis model comprises a multilingual speech synthesis model; and

the operations further comprise pre-training the encoder of the speech synthesis model using a classification objective to predict a classification token of the plurality of tokens of the input encoder embedding, the classification token comprising a language identifier.

8. The method of claim 1 , wherein:

the speech synthesis model comprises a multilingual speech synthesis model; and

the output encoder embedding comprises a sequence of encoder tokens, each encoder token comprising language information about the input text.

9. The method of claim 1 , wherein:

the speech synthesis model comprises a multi-accent speech synthesis model; and

the operations further comprise pre-training the encoder of the speech synthesis model using a classification objective to predict a classification token of the plurality of tokens of the input encoder embedding, the classification token comprising an accent identifier.

10. 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, at an encoder of a speech synthesis model, a text input comprising a sequence of words represented as an input encoder embedding, the input encoder embedding comprising a plurality of tokens, the plurality of tokens comprising a first set of grapheme tokens representing the text input as respective graphemes and a second set of phoneme tokens representing the text input as respective phonemes, each grapheme token of the first set of grapheme tokens comprising a respective wordpiece sub-word of a respective word in the sequence of words, wherein each corresponding token of the plurality of tokens of the input encoder embedding represents a combination of:

a respective word position embedding for each respective word in the sequence of words, the respective word position embedding representing sub-word level positions for both one or more of the grapheme tokens from the first set of grapheme tokens that correspond to the respective word and one or more of the phoneme tokens from the second set of phoneme tokens that correspond to the respective word; and

a position embedding representing an overall index of position for each token of the plurality of tokens of the input encoder embedding;

for each respective phoneme token of the second set of phoneme tokens:

identifying, by the encoder, a respective word of the sequence of words corresponding to the respective phoneme token based on the respective word position embedding that represents the sub-word level position for the respective phoneme token that corresponds to the respective word; and

determining, by the encoder, a respective grapheme token representing the respective word of the sequence of words corresponding to the respective phoneme token by determining that the sub-word level position for the respective grapheme token that corresponds to the respective word is represented by the same respective word position embedding as the respective word position embedding representing the sub-word level position for the respective phoneme token; and

generating, by the encoder, an output encoder embedding based on a relationship between each respective phoneme token and the respective grapheme token determined to represent a same respective word as the respective phoneme token.

11. The system of claim 10 , wherein the combination representing each token of the plurality of tokens of the input encoder embedding further comprises:

one of a grapheme token embedding or a phoneme token embedding; and

a segment embedding.

12. The system of claim 10 , wherein the speech synthesis model comprises an attention mechanism in communication with the encoder.

13. The system of claim 10 , wherein the speech synthesis model comprises a duration-based upsampler in communication with the encoder.

14. The system of claim 10 , wherein the plurality of tokens of the input encoder embedding comprises a special token identifying a language of the input text.

15. The system of claim 10 , wherein the operations further comprise:

pre-training the encoder of the speech synthesis model by:

feeding the encoder a plurality of training examples, each training example represented as a sequence of training grapheme tokens corresponding to a training sequence of words and a sequence of training phoneme tokens corresponding to the same training sequence of words;

masking a training phoneme token from the sequence of training phoneme tokens for a respective word from the training sequence of words; and

masking a training grapheme token from the sequence of training phoneme tokens for the respective word from the training sequence of words.

16. The system of claim 10 , wherein:

the speech synthesis model comprises a multilingual speech synthesis model; and

the operations further comprise pre-training the encoder of the speech synthesis model using a classification objective to predict a classification token of the plurality of tokens of the input encoder embedding, the classification token comprising a language identifier.

17. The system of claim 10 , wherein:

the speech synthesis model comprises a multilingual speech synthesis model; and

the output encoder embedding comprises a sequence of encoder tokens, each encoder token comprising language information about the input text.

18. The system of claim 10 , wherein:

the speech synthesis model comprises a multi-accent speech synthesis model; and

the operations further comprise pre-training the encoder of the speech synthesis model using a classification objective to predict a classification token of the plurality of tokens of the input encoder embedding, the classification token comprising an accent identifier.