Sample-efficient adaptive text-to-speech

1. A method performed by one or more computers, the method comprising:

generating an adaptive audio-generation model including learning a plurality of embedding vectors and parameter values of a neural network using training data comprising first text and audio data representing a plurality of different individual speakers speaking portions of the first text, wherein the plurality of embedding vectors represent respective voice characteristics of the plurality of different individual speakers; and

adapting the adaptive audio-generation model for a new individual speaker using adaptation data comprising second text and audio data representing the new individual speaker speaking portions of the second text, the new individual speaker being different from each of the plurality of individual speakers, wherein adapting the audio-generation model includes (i) learning values of a new embedding vector for the new individual speaker and (ii) adjusting the parameter values of the neural network that were learned during the generation of the adaptive audio-generation model.

2. The method of claim 1 , wherein the audio-generation model adapted for the new individual speaker does not use any of the plurality of embedding vectors for the plurality of individual speakers in the training data.

3. The method of claim 1 , wherein generating the adaptive audio-generation model comprises learning a different respective embedding vector for each of the plurality of individual speakers.

4. The method of claim 3 , wherein adapting the adaptive audio-generation model comprises extracting features from the adaptation data and learning values of the new embedding vector that are most likely according to the learned parameter values of the audio-generation model.

5. The method of claim 3 , wherein adapting the adaptive audio-generation model comprises extracting features from the adaptation data and using at least the extracted features to learning both values of the new embedding vector and new parameter values for the neural network.

6. The method of claim 5 , further comprising:

pre-initializing the new embedding vector with values learned from the features of the adaptation data.

7. The method of claim 1 , wherein generating the adaptive audio-generation model comprises learning an encoder network that is configured to predict an embedding vector for a speaker when given adaptation data for the speaker.

8. The method of claim 7 , wherein the encoder network comprises a summation of two subnetworks comprising 1) a pre-trained speaker verification model, and 2) a second sub-network comprising a plurality of convolutional layers configured to learn residual features of individual speakers.

9. The method of claim 1 , further comprising:

performing text-to-speech generation using the audio-generation model adapted for the new individual speaker, including using as input the learned new embedding vector and features of a new portion of text.

10. The method of claim 9 , wherein performing text-to-speech generation comprises automatically generating audio of a translation of video or audio content, the audio of the translation being adapted to the characteristics of the new speaker.

11. The method of claim 9 , wherein performing text-to-speech generation comprises automatically generating audio of a translation of speech during a video conference or a phone call, the audio of the translation being adapted to the characteristics of the new speaker.

12. The method of claim 1 , wherein the training data comprises multiple hours of audio recordings for each individual speaker of the plurality of different individual speakers, and wherein the adaptation data comprises less than ten minutes of audio recordings of the new individual speaker.

13. The method of claim 1 , wherein generating the adaptive audio-generation model comprises training the model in a datacenter, and

wherein adapting the audio-generation model for the new individual speaker comprises adapting the model on a mobile device or a single internet-enabled device.

14. A system comprising:

one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:

generating an adaptive audio-generation model including learning a plurality of embedding vectors and parameter values of a neural network using training data comprising first text and audio data representing a plurality of different individual speakers speaking portions of the first text, wherein the plurality of embedding vectors represent respective voice characteristics of the plurality of different individual speakers; and

adapting the adaptive audio-generation model for a new individual speaker using adaptation data comprising second text and audio data representing the new individual speaker speaking portions of the second text, the new individual speaker being different from each of the plurality of individual speakers, wherein adapting the audio-generation model includes (i) learning values of a new embedding vector for the new individual speaker and (ii) adjusting the parameter values of the neural network that were learned during the generation of the adaptive audio-generation model.

15. The system of claim 14 , wherein the audio-generation model adapted for the new individual speaker does not use any of the plurality of embedding vectors for the plurality of individual speakers in the training data.

16. The system of claim 14 , wherein generating the adaptive audio-generation model comprises learning a different respective embedding vector for each of the plurality of individual speakers.

17. The system of claim 16 , wherein adapting the adaptive audio-generation model comprises extracting features from the adaptation data and learning values of the new embedding vector that are most likely according to the learned parameter values of the audio-generation model.

18. The system of claim 16 , wherein adapting the adaptive audio-generation model comprises extracting features from the adaptation data and using at least the extracted features to learn both values of the new embedding vector and new parameter values for the neural network.

19. The system of claim 18 , wherein the operations further comprise:

pre-initializing the new embedding vector with values learned from the features of the adaptation data.

20. One or more non-transitory computer storage media encoded with computer program instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

generating an adaptive audio-generation model including learning a plurality of embedding vectors and parameter values of a neural network using training data comprising first text and audio data representing a plurality of different individual speakers speaking portions of the first text, wherein the plurality of embedding vectors represent respective voice characteristics of the plurality of different individual speakers; and

adapting the adaptive audio-generation model for a new individual speaker using adaptation data comprising second text and audio data representing the new individual speaker speaking portions of the second text, the new individual speaker being different from each of the plurality of individual speakers, wherein adapting the audio-generation model includes (i) learning values of a new embedding vector for the new individual speaker and (ii) adjusting the parameter values of the neural network that were learned during the generation of the adaptive audio-generation model.