Low-power multi-voice assistants voice activation
Systems and methods presented herein generally include multi-wake phrase detection executed on a single device utilizing multiple voice assistants. Systems and methods presented herein can further include continuously running a Voice Activity Detection (VAD) process which detects presence of human speech. The multi-wake phrase detection can activate when the VAD process detects human speech. Once activated, the multi-wake phrase detection can determine which (if any) of the wake phrases of the multiple voice assistants might be in the detected speech. Operation of the multi-wake phrase detection can have a low miss-rate. In some examples, operation of the multi-wake phrase detection can be granular to accomplish the low miss-rates at low power with a tolerance for false positives on wake phrase detection.
1. A multi-wake word/phrase detector configured to:
activate upon detection of human speech in an audio signal;
receive the audio signal;
analyze the audio signal to identify a presence of one specific wake word/phrase from a plurality of potential wake words/phrases;
activate an on-device processor configured to execute a vendor-specific wake word/phrase detection algorithm upon identifying the presence of the one specific wake word/phrase for the vendor, without a simultaneous activation of wake word/phrase detection algorithms for other vendors unassociated with the specific wake word/phrase; and
transmit an interrupt to activate an application processor to initiate streaming of the audio signal to a remote processor for further processing.
2. The multi-wake word/phrase detector of claim 1 further configured to return to an idle mode after the audio signal and the indicator are transmitted.
3. The multi-wake word/phrase detector of claim 1 , wherein the analysis of the audio signal is performed by a multi-class classification neural network trained to recognize the presence of the specific wake word/phrase from the plurality of potential wake words/phrases.
4. The multi-wake word/phrase detector of claim 3 , wherein the multi-class classification neural network includes a one-to-one mapping of a classification to a given one of the plurality of potential wake words/phrases for which the multi-class classification neural network is trained.
5. The multi-wake word/phrase detector of claim 4 , wherein the classification corresponds to one of a plurality of voice assistants.
6. The multi-wake word/phrase detector of claim 3 , wherein the multi-class classification neural network has a fixed topology.
7. A system comprising:
a voice activity detector configured to transmit audio signals including human speech;
a multi-wake word/phrase detector configured to receive the audio signals from the voice activity detector, analyze the audio signal to identify a presence of one specific wake word/phrase from a plurality of potential wake words/phrases;
a multi-wake word/phrase detection engine configured to receive the audio signal from the multi-wake word/phrase detector, to execute, on the audio signal, a vendor-specific wake word/phrase detection algorithm upon identifying the presence of the one specific wake word/phrase for the vendor indicated by the multi-wake word/phrase detector, without a simultaneous activation of wake word/phrase detection algorithms for other vendors unassociated with the specific wake word/phrase
an application processor configured to activate in response to an interrupt from the multi-wake word/phrase detection engine upon the identification of the specific wake word/phrase and initiated streaming of the audio signal to a remote processor for natural language recognition.
8. The system of claim 7 , wherein the multi-wake word/phrase detector is implemented with a multi-class classification neural network including a one-to-one mapping of a classification to a given one of the plurality of potential wake words/phrases for which the multi-class classification neural network is trained.
9. The system of claim 8 , wherein the multi-class classification neural network has a fixed topology.
10. The system of claim 7 , wherein the voice activity detector maintains at least the multi-wake word/phrase detector, the multi-wake word/phrase detection engine, and the application processor in an idle mode.
11. The system of claim 7 , wherein the multi-wake word/phrase detection engine includes an on-device processor configured to execute one of multiple detection algorithms each specific to a voice assistant.
12. The system of claim 11 , further comprising:
a central listening device incorporating the voice activity detector, and the multi-wake word/phrase detector.
13. The system of claim 7 , further comprising:
one or more peripheral voice activated devices each including a dedicated wake word/phrase detection processor in communication with the multi-wake word/phrase detector, each of the peripheral voice activated devices being associated with a specific one of the wake word/phrase detection algorithms.
14. The system of claim 13 , wherein at least one of the peripheral voice activated devices is battery-powered.
15. The system of claim 7 , wherein the application processor transmits the audio signal to a cloud processing system.
16. A method comprising:
detecting human speech in an audio signal;
analyzing the audio signal to identify a presence of one specific wake word/phrase from a plurality of potential wake words/phrases;
activating an on-device processor configured to execute a vendor-specific wake word/phrase detection algorithm upon identifying the presence of the one specific wake word/phrase for the vendor, without a simultaneous activation of wake word/phrase detection algorithms for other vendors unassociated with the specific wake word/phrase; and
generating an interrupt to activate an application processor to initiate streaming of the audio signal to a remote processor for further processing.
17. The method of claim 16 , further comprising:
executing, on the audio signal, only wake word/phrase detection algorithm(s) associated with the one or more wake words/phrases indicated by the indicator;
identifying a spoken wake word/phrase of the plurality of wake words/phrases;
activating an application processor upon the identification of the spoken wake word/phrase; and
transmitting the audio signal for natural language recognition.
18. The method of claim 16 , wherein the categorizing of the audio signal is performed by a multi-class classification neural network trained to recognize the potential presence of one or more wake words/phrases.
19. The method of claim 18 , wherein the multi-class classification neural network includes a one-to-one mapping of a classification to a given one of the plurality of potential wake words/phrases for which the multi-class classification neural network is trained.
20. The method of claim 18 , wherein the multi-class classification neural network has a fixed topology.