Predicting spectral representations for training speech synthesis neural networks
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network to perform speech synthesis. One of the methods includes obtaining a training data set for training a first neural network to process a spectral representation of an audio sample and to generate a prediction of the audio sample, wherein, after training, the first neural network obtains spectral representations of audio samples from a second neural network; for a plurality of audio samples in the training data set: generating a ground-truth spectral representation of the audio sample; and processing the ground-truth spectral representation using a third neural network to generate an updated spectral representation of the audio sample; and training the first neural network using the updated spectral representations, wherein the third neural network is configured to generate updated spectral representations that resemble spectral representations generated by the second neural network.
1. A method of training a first neural network to process a spectral representation of an audio sample and to generate a prediction of the audio sample,
wherein, after the first neural network has been trained, the first neural network is configured to obtain the spectral representation of an audio sample to be processed from a second neural network,
wherein the second neural network is configured to process a text input and to generate a predicted spectral representation of a vocalization of the text input,
the method comprising:
obtaining a training data set comprising a plurality of audio samples;
for one or more audio samples in the training data set:
generating a ground-truth spectral representation of the audio sample; and
processing the ground-truth spectral representation of the audio sample using a third neural network to generate an updated spectral representation of the audio sample; and
training the first neural network using the updated spectral representations,
wherein the third neural network is configured to process ground-truth spectral representations and to generate updated spectral representations that resemble spectral representations generated by the second neural network.
2. The method of claim 1 , wherein the third neural network is a pretrained autoencoder neural network that has been trained to perform operations comprising:
obtaining the ground-truth spectral representation of an input audio sample;
processing the ground-truth spectral representation using an encoder subnetwork to generate an embedding of the ground-truth spectral representation; and
processing embedding of the ground-truth spectral representation using a decoder subnetwork to generate the updated spectral representation of the input audio sample.
3. The method of claim 1 , wherein the first neural network is configured to generate audio samples characterizing any voice.
4. The method of claim 3 , wherein the second neural network is only configured to generate predicted spectral representations corresponding to a predetermined set of voices.
5. The method of claim 1 , wherein the third neural network has been trained by performing operations comprising:
obtaining a plurality of training examples that each comprise i) a training ground-truth spectral representation of a training audio sample and ii) a training predicted spectral representation of the training audio sample generated by the second neural network;
for one or more of the plurality of training examples:
processing the training ground-truth spectral representation using the third neural network to generate a training updated spectral representation of the training audio sample; and
determining an error between the training updated spectral representation and the training predicted spectral representation; and
generating a parameter update for the third neural network using the determined errors.
6. The method of claim 1 , wherein training the first neural network using the updated spectral representations comprises:
training the first neural network and the second neural network jointly; and
finetuning the first neural network using the updated spectral representations generated by the third neural network.
7. The method of claim 1 , further comprising deploying the first neural network and the second neural network in a production setting without the third neural network.
8. A method comprising processing a spectral representation of an audio sample using a first neural network to generate a prediction of the audio sample,
wherein the first neural network is configured to obtain the spectral representation of the audio sample from a second neural network,
wherein the second neural network is configured to process a text input and to generate a predicted spectral representation of a vocalization of the text input,
and wherein the first neural network has been trained by performing operations comprising:
obtaining a training data set comprising a plurality of audio samples;
for one or more audio samples in the training data set:
generating a ground-truth spectral representation of the audio sample; and
processing the ground-truth spectral representation of the audio sample using a third neural network to generate an updated spectral representation of the audio sample; and
training the first neural network using the updated spectral representations,
wherein the third neural network is configured to process ground-truth spectral representations and to generate updated spectral representations that resemble spectral representations generated by the second neural network.
9. The method of claim 8 , wherein the third neural network is a pretrained autoencoder neural network that has been trained to perform operations comprising:
obtaining the ground-truth spectral representation of an input audio sample;
processing the ground-truth spectral representation using an encoder subnetwork to generate an embedding of the ground-truth spectral representation; and
processing embedding of the ground-truth spectral representation using a decoder subnetwork to generate the updated spectral representation of the input audio sample.
10. The method of claim 8 , wherein the first neural network is configured to generate audio samples characterizing any voice.
11. The method of claim 10 , wherein the second neural network is only configured to generate predicted spectral representations corresponding to a predetermined set of voices.
12. The method of claim 8 , wherein the third neural network has been trained by performing operations comprising:
obtaining a plurality of training examples that each comprise i) a training ground-truth spectral representation of a training audio sample and ii) a training predicted spectral representation of the training audio sample generated by the second neural network;
for one or more of the plurality of training examples:
processing the training ground-truth spectral representation using the third neural network to generate a training updated spectral representation of the training audio sample; and
determining an error between the training updated spectral representation and the training predicted spectral representation; and
generating a parameter update for the third neural network using the determined errors.
13. The method of claim 8 , wherein training the first neural network using the updated spectral representations comprises:
training the first neural network and the second neural network jointly; and
finetuning the first neural network using the updated spectral representations generated by the third neural network.
14. The method of claim 8 , wherein the first neural network and the second neural network have been deployed in a production setting without the third neural network.
15. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations for training a first neural network to process a spectral representation of an audio sample and to generate a prediction of the audio sample,
wherein, after the first neural network has been trained, the first neural network is configured to obtain the spectral representation of an audio sample to be processed from a second neural network,
wherein the second neural network is configured to process a text input and to generate a predicted spectral representation of a vocalization of the text input,
the operations comprising:
obtaining a training data set comprising a plurality of audio samples;
for one or more audio samples in the training data set:
generating a ground-truth spectral representation of the audio sample; and
processing the ground-truth spectral representation of the audio sample using a third neural network to generate an updated spectral representation of the audio sample; and
training the first neural network using the updated spectral representations,
wherein the third neural network is configured to process ground-truth spectral representations and to generate updated spectral representations that resemble spectral representations generated by the second neural network.
16. The system of claim 15 , wherein the third neural network is a pretrained autoencoder neural network that has been trained to perform operations comprising:
obtaining the ground-truth spectral representation of an input audio sample;
processing the ground-truth spectral representation using an encoder subnetwork to generate an embedding of the ground-truth spectral representation; and
processing embedding of the ground-truth spectral representation using a decoder subnetwork to generate the updated spectral representation of the input audio sample.
17. The system of claim 15 , wherein the first neural network is configured to generate audio samples characterizing any voice.
18. The system of claim 17 , wherein the second neural network is only configured to generate predicted spectral representations corresponding to a predetermined set of voices.
19. The system of claim 15 , wherein the third neural network has been trained by performing operations comprising:
obtaining a plurality of training examples that each comprise i) a training ground-truth spectral representation of a training audio sample and ii) a training predicted spectral representation of the training audio sample generated by the second neural network;
for one or more of the plurality of training examples:
processing the training ground-truth spectral representation using the third neural network to generate a training updated spectral representation of the training audio sample; and
determining an error between the training updated spectral representation and the training predicted spectral representation; and
generating a parameter update for the third neural network using the determined errors.
20. The system of claim 15 , wherein training the first neural network using the updated spectral representations comprises:
training the first neural network and the second neural network jointly; and
finetuning the first neural network using the updated spectral representations generated by the third neural network.