Resolving time-delays using generative models
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating audio output samples predicted to be communicated by a user. One example system includes a first user device having a first user. The first user device initiates a communication session between the first user and a second user of a second user device. The first user device obtains a neural network model of the second user. The neural network model is trained to generate, conditioned on audio input samples received up to a current time step, an audio output sample predicted to be communicated by the second user at a next time step. The user device repeatedly provides received audio input samples as input to the neural network model and plays audio output samples generated by the neural network model in place of received audio input samples communicated by the second user.
1. A method comprising:
initiating a communication session during which a first user device transmits audio data to and receives audio data from a second user device; and
at each of multiple time steps during the communication session:
providing a model input comprising past audio samples received by the first user device up to a current time step to a neural network model,
generating, by using the neural network model and from the model input, a predicted audio sample that is a prediction of a current audio sample to be communicated by the second user at the current time step, and
playing the predicted audio sample generated by using the neural network model before the current audio sample communicated by the second user device at the current time step is ready to be played.
2. The method of claim 1 , wherein the model input to the neural network model further comprises past audio samples generated by the first user device up to the current time step.
3. The method of claim 2 , wherein the model input to the neural network model further comprises a current audio sample generated by the first user device at the current time step.
4. The method of claim 1 , wherein playing the predicted audio sample generated by using the neural network model at the current time step comprises:
playing the predicted audio sample before the current audio sample is received by the first user device from the second user device.
5. The method of claim 1 , wherein the neural network model is configured to generate the predicted audio sample conditioned on a context vector for the second user, wherein the context vector defines a context of the communications between the first user and the second user, and wherein the method further comprises obtaining the context vector for the second user and conditioning the neural network model on the context vector.
6. The method of claim 1 , wherein generating the predicted audio sample comprises:
generating a probability distribution over a set of possible predicted audio samples for the current time step; and
selecting a predicted audio sample corresponding to a highest probability in the probability distribution.
7. The method of claim 6 , wherein the set of possible predicted audio samples includes one or more samples representing silence.
8. The method of claim 6 , wherein the set of possible predicted audio samples includes one or more samples representing phonemes.
9. The method of claim 6 , wherein the method further comprises, at a particular time step during the communication session:
providing a model input comprising past audio samples received by the first user device up to the particular time step to the neural network model;
generating a probability distribution over the set of possible predicted audio samples for the particular time step;
determining that none of the probabilities in the probability distribution meet a predetermined threshold; and
in response, playing a current audio sample communicated by the second user device at the particular time step instead of any predicted audio sample generated by using the neural network model.
10. 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 comprising:
initiating a communication session during which a first user device transmits audio data to and receives audio data from a second user device; and
at each of multiple time steps during the communication session:
providing a model input comprising past audio samples received by the first user device up to a current time step to a neural network model,
generating, by using the neural network model and from the model input, a predicted audio sample that is a prediction of a current audio sample to be communicated by the second user at the current time step, and
playing the predicted audio sample generated by using the neural network model before the current audio sample communicated by the second user device at the current time step is ready to be played.
11. The system of claim 10 , wherein the model input to the neural network model further comprises past audio samples generated by the first user device up to the current time step.
12. The system of claim 11 , wherein the model input to the neural network model further comprises a current audio sample generated by the first user device at the current time step.
13. The system of claim 10 , wherein playing the predicted audio sample generated by using the neural network model at the current time step comprises:
playing the predicted audio sample before the current audio sample is received by the first user device from the second user device.
14. The system of claim 10 , wherein the neural network model is configured to generate the predicted audio sample conditioned on a context vector for the second user, wherein the context vector defines a context of the communications between the first user and the second user, and wherein the method further comprises obtaining the context vector for the second user and conditioning the neural network model on the context vector.
15. The system of claim 10 , wherein generating the predicted audio sample comprises:
generating a probability distribution over a set of possible predicted audio samples for the current time step; and
selecting a predicted audio sample corresponding to a highest probability in the probability distribution.
16. The system of claim 15 , wherein the set of possible predicted audio samples includes one or more samples representing silence.
17. The system of claim 15 , wherein the set of possible predicted audio samples includes one or more samples representing phonemes.
18. The system of claim 15 , wherein the operations further comprise, at a particular time step during the communication session:
providing a model input comprising past audio samples received by the first user device up to the particular time step to the neural network model;
generating a probability distribution over the set of possible predicted audio samples for the particular time step;
determining that none of the probabilities in the probability distribution meet a predetermined threshold; and
in response, playing a current audio sample communicated by the second user device at the particular time step instead of any predicted audio sample generated by using the neural network model.
19. One or more non-transitory computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
initiating a communication session during which a first user device transmits audio data to and receives audio data from a second user device; and
at each of multiple time steps during the communication session:
providing a model input comprising past audio samples received by the first user device up to a current time step to a neural network model,
generating, by using the neural network model and from the model input, a predicted audio sample that is a prediction of a current audio sample to be communicated by the second user at the current time step, and
playing the predicted audio sample generated by using the neural network model before the current audio sample communicated by the second user device at the current time step is ready to be played.