Detecting machine-outputted audio

1. A computer-implemented method comprising:

receiving first audio data;

determining first audio embedding data using a first machine learning (ML) model and a portion of the first audio data;

determining, using the first audio embedding data, that the first audio data corresponds to machine-outputted audio;

in response to determining that the first audio data corresponds to machine-outputted audio, storing the first audio data;

receiving second audio data;

determining, using the first ML model and the second audio data, second audio embedding data;

determining, using the second audio embedding data, that the second audio data corresponds to non-machine-outputted audio;

in response to determining that the second audio data corresponds to non-machine-outputted audio, determining the second audio data includes a user input relating to content of the first audio data; and

based at least in part on the first audio data, determining first output data responsive to the user input.

2. The computer-implemented method of claim 1 , further comprising:

determining that the second audio data corresponds to non-machine-outputted audio; and

in response to determining that the second audio data corresponds to non-machine-outputted audio, sending the second audio data to an automatic speech recognition (ASR) component for processing.

3. The computer-implemented method of claim 1 , further comprising:

processing, using an automatic speech recognition (ASR) component, the first audio data; and

in response to determining that the first audio data corresponds to machine-outputted audio, ceasing processing by the ASR component.

4. The computer-implemented method of claim 1 , further comprising:

receiving first audio;

determining that the first audio includes speech; and

in response to determining that the first audio includes speech, processing the first audio to determine the first audio data, wherein the first audio data includes log Mel-filterbank energies feature data.

5. The computer-implemented method of claim 1 , wherein the first audio data is received at a device, and the method further comprises:

in response to determining that the first audio data corresponds to machine-outputted audio, sending, to the device, second output data indicating the first audio data was outputted by a machine.

6. The computer-implemented method of claim 1 , further comprising:

determining first automatic speech recognition (ASR) data corresponding to the first audio data;

determining second ASR data corresponding to the second audio data; and

determining, using the second ASR data, natural language understanding (NLU) data corresponding to the user input,

wherein determining that the second audio data includes the user input relating to content of the first audio data is based on processing the first ASR data and the NLU data.

7. The computer-implemented method of claim 1 ,

wherein determining that the first audio data corresponds to machine-outputted audio further comprises determining that the first audio data corresponds to a synthesized speech category.

8. The computer-implemented method of claim 1 , further comprising:

in response to determining that the first audio data corresponds to machine outputted audio, generating second output data indicating the first audio data was outputted by a machine; and

sending the second output data to a system component.

9. The computer-implemented method of claim 1 ,

wherein determining that the first audio data corresponds to machine-outputted audio further comprises determining that the first audio data corresponds to a recorded audio category, and

wherein determining that the second audio data includes the user input relating to the content of the first audio data is based on determining that the first audio data corresponds to the recorded audio category.

10. A system comprising:

at least one processor; and

at least one memory including instructions that, when executed by the at least one processor, cause the system to:

receive first audio data;

determine first audio embedding data using a first machine learning (ML) model and a portion of the first audio data;

determine, using the first audio embedding data, that the first audio data corresponds to machine-outputted audio;

in response to determining that the first audio data corresponds to machine-outputted audio, storing the first audio data;

receive second audio data;

determine, using the first ML model and the second audio data, second audio embedding data;

determining, using the second audio embedding data, that the second audio data corresponds to non-machine-outputted audio;

in response to determining that the second audio data corresponds to non-machine-outputted audio, determine the second audio data includes a user input relating to content of the first audio data; and

based at least in part on the first audio data, determining second output data responsive to the user input.

11. The system of claim 10 , wherein the at least one memory includes further instructions that, when executed by the at least one processor, further cause the system to:

determine that the second audio data corresponds to non-machine-outputted audio; and

in response to determining that the second audio data corresponds to non-machine-outputted audio, send the second audio data to an automatic speech recognition (ASR) component for processing.

12. The system of claim 10 , wherein the at least one memory includes further instructions that, when executed by the at least one processor, further cause the system to:

process, using an automatic speech recognition (ASR) component, the first audio data; and

in response to determining that the first audio data corresponds to machine-outputted audio, cease processing by the ASR component.

13. The system of claim 10 , wherein the at least one memory includes further instructions that, when executed by the at least one processor, further cause the system to:

receive first audio;

determine that the first audio includes speech; and

in response to determining that the first audio includes speech, process the first audio to determine the first audio data, wherein the first audio data includes log Mel-filterbank energies feature data.

14. The system of claim 10 , wherein the first audio data is received at a device, and wherein the at least one memory includes further instructions that, when executed by the at least one processor, further cause the system to:

in response to determining that the first audio data corresponds to machine-outputted audio, send, to the device, second output data indicating the first audio data was outputted by a machine.

15. The system of claim 10 , wherein the at least one memory includes further instructions that, when executed by the at least one processor, further cause the system to:

determine first automatic speech recognition (ASR) data corresponding to the first audio data;

determine second ASR data corresponding to the second audio data; and

determine, using the second ASR data, natural language understanding (NLU) data corresponding to the user input;

wherein determining that the second audio data includes the user input relating to content of the first audio data is based on processing the first ASR data and the NLU data.

16. The system of claim 10 ,

wherein the instruction that causes the system to determine that the first audio data corresponds to machine-outputted audio further causes the system to determine that the first audio data corresponds to a synthesized speech category.

17. The system of claim 10 , wherein the at least one memory includes further instructions that, when executed by the at least one processor, further cause the system to:

in response to determining that the first audio data corresponds to machine outputted audio, generate second output data indicating the first audio data was outputted by a machine; and

send the second output data to a system component.

18. The system of claim 10 ,

Wherein the instruction that causes the system to determine that the first audio data corresponds to machine-outputted audio further causes the system to determine that the first audio data corresponds to a recorded audio category, and

wherein determining that the second audio data includes the user input relating to the content of the first audio data is based on determining that the first audio data corresponds to the recorded audio category.

19. A computer-implemented method comprising:

receiving first audio;

determining, using a Long Short-Term Memory (LSTM) model and a portion of the first audio, first audio embedding data corresponding to the first audio;

processing the first audio embedding data using a logistic regression model to determine that the first audio corresponds to machine-outputted audio;

in response to determining that the first audio data corresponds to machine outputted audio, storing the first audio data;

after receiving the first audio, receiving, by the device, second audio;

determining, using the LSTM model and a portion of the second audio, second audio embedding data corresponding to the second audio;

determining, using the second audio embedding data, that the second audio corresponds to non-machine-outputted audio;

in response to determining that the second audio corresponds to non-machine-outputted audio, processing the second audio embedding data using an automatic speech recognition (ASR) component to determine first ASR data corresponding to a spoken user input represented in the second audio;

determining, using the first ASR data and content represented in the first audio, that the spoken user input relates to the first audio;

determining, using the first ASR data and the first audio, first output data responsive to the spoken user input; and

causing presentation of the first output data.

20. The computer-implemented method of claim 19 , further comprising:

wherein determining the second audio corresponds to non-machine-outputted audio is based on processing the second audio embedding data using the logistic regression model.

21. The computer-implemented method of claim 19 , further comprising:

processing the first audio to determine that the first audio includes speech;

in response to determining that the first audio includes speech, processing the first audio to determine that the first audio includes a representation of a wakeword; and

in response to the first audio including the representation of the wakeword, determining first feature data corresponding to the first audio.