Systems and methods for power-efficient keyword detection

1. A network microphone device (NMD) comprising:

a plurality of microphones;

a network interface;

one or more processors; and

a tangible, non-transitory, computer-readable medium storing instructions that, when executed by the one or more processors, cause the NMD to perform operations comprising:

detecting sound via at least one of the microphones;

determining, via a voice activity detection process, whether the detected sound includes voice activity;

capturing the detected sound as sound data;

while in a first stage, forgoing spatial processing of the captured sound data;

transitioning the NMD from the first stage to a second stage after determining that the detected sound includes voice activity; and

while in the second stage:

spatially processing the captured sound data using a spatial processor to produce filtered sound data;

detecting, via a wake-word engine, a wake word based on captured the filtered sound data; and

after detecting the wake word, determining an action to be performed based on the captured sound data.

2. The NMD of claim 1 , wherein the captured sound data comprises less than about 300 milliseconds of sound data.

3. The NMD of claim 1 , wherein determining the action to be performed based on the captured sound data comprises:

transmitting, via the network interface, at least a portion of the captured sound data to one or more remote computing devices associated with a voice assistant service corresponding to the detected wake word for processing a voice utterance.

4. The NMD of claim 1 , wherein the operations further comprise performing the determined action, wherein the action comprises controlling playback of audio content.

5. The NMD of claim 1 , wherein, in the first stage, the one or more processors operate at a lower average clock speed than in the second stage.

6. The NMD of claim 1 , wherein, in the first stage, the NMD consumes less power than in the second stage.

7. The NMD of claim 1 , wherein, in the first stage, the spatial processor is in a low-power standby mode.

8. The NMD of claim 1 , wherein the operations further comprise:

after determining the action to be performed, transitioning the NMD from the second stage to the first stage.

9. A method comprising:

detecting sound via at least one microphone of a network microphone device (NMD);

determining, via a voice activity detection process of the NMD, whether the detected sound includes voice activity;

capturing the detected sound as sound data;

while in a first stage of the NMD, forgoing spatial processing of the captured sound data;

transitioning the NMD from the first stage to a second stage after determining that the detected sound includes voice activity; and

while in the second stage:

spatially processing the captured sound data using a spatial processor to produce filtered sound data;

detecting, via a wake-word engine, a wake word based on the filtered sound data; and

after detecting the wake word, determining an action to be performed based on the captured sound data.

10. The method of claim 9 , wherein the captured sound data comprises less than about 300 milliseconds of sound data.

11. The method of claim 9 , wherein determining the action to be performed based on the captured sound data comprises:

transmitting, via a network interface of the NMD, at least a portion of the captured sound data to one or more remote computing devices associated with a voice assistant service corresponding to the detected wake word for processing a voice utterance.

12. The method of claim 9 , further comprising performing the determined action, wherein the action comprises controlling playback of audio content.

13. The method of claim 9 , wherein, in the first stage, the NMD consumes less power than in the second stage.

14. The method of claim 9 , further comprising, after determining the action to be performed, transitioning the NMD from the second stage to the first stage.

15. A tangible, non-transitory computer-readable medium storing instructions that, when executed by one or more processors of a network microphone device (NMD), cause the NMD to perform operations comprising:

detecting sound via at least one microphone of the NMD;

determining, via a voice activity detection process, whether the detected sound includes voice activity;

capturing the detected sound as sound data;

while in a first stage, forgoing spatial processing of the captured sound data;

transitioning the NMD from the first stage to a second stage after determining that the detected sound includes voice activity; and

while in the second stage:

spatially processing the captured sound data using a spatial processor to produce filtered sound data;

detecting, via a wake-word engine, a wake word based on the filtered sound data; and

after detecting the wake word, determining an action to be performed based on the captured sound data.

16. The computer-readable medium of claim 15 , wherein the captured sound data comprises less than about 300 milliseconds of sound data.

17. The computer-readable medium of claim 15 , wherein determining the action to be performed based on the captured sound data comprises:

transmitting, via a network interface of the NMD, at least a portion of the captured sound data to one or more remote computing devices associated with a voice assistant service corresponding to the detected wake word for processing a voice utterance.

18. The computer-readable medium of claim 15 , wherein the operations further comprise performing the determined action, wherein the action comprises controlling playback of audio content.

19. The computer-readable medium of claim 15 , wherein, in the first stage, the NMD consumes less power than in the second stage.

20. The computer-readable medium of claim 15 , wherein the operations further comprise:

after determining the action to be performed, transitioning the NMD from the second stage to the first stage.