Prediction based wake-word detection and methods for use therewith
An audio processing circuit includes an audio front end (AFE) configured to generate audio data in response to audio signals received from one or more microphones. One or more special purpose engines (SPEs) are configured to: generate prediction data via a usage model; generate bias adjustment data based on the prediction data; adjust a threshold of an audio event recognition engine based on the prediction data; and determine, via the audio event recognition engine and based on audio data, when an audio event is detected.
1. An audio processing circuit of a client device comprising:
an audio front end (AFE) of the audio processing circuit configured to generate audio data in response to audio signals received from one or more microphones;
one or more special purpose engines (SPEs) of the audio processing circuit configured to:
generate prediction data for a current time via a an always-on voice (AOV) usage model, wherein the AOV usage model is utilized as a function of time of day;
generate bias adjustment data corresponding to the current time of day, based on the prediction data;
adjust a threshold of an audio event recognition engine of the audio processing circuit for the current time of day, based on the prediction data, wherein the audio event recognition engine is a wake-word recognition engine; and
determine, via the audio event recognition engine and based on audio data, when an audio event is detected based on the generated bias adjustment data and the adjusted threshold.
2. The audio processing circuit of claim 1 , wherein the prediction data includes always-on voice AOV prediction data, wherein the AOV usage model is generated based on historical data of AOV usage, wherein the threshold is a wake-word threshold and wherein the audio event corresponds to detection of a wake-word.
3. The audio processing circuit of claim 2 , wherein the historical data of AOV usage is based on times of wake-word detections.
4. The audio processing circuit of claim 3 , wherein the one or more SPEs further includes a query recognition engine that generates recognized query data based on the audio data after the wake-word is detected.
5. The audio processing circuit of claim 4 , wherein the historical data of AOV usage is further based on the recognized query data.
6. The audio processing circuit of claim 5 , wherein the wake-words are only considered in the historical data of AOV when followed by the recognized query data.
7. The audio processing circuit of claim 2 , wherein the wake-word recognition engine includes a model anchored by two sets of posterior parameters, only differing in threshold.
8. The audio processing circuit of claim 7 , wherein the two sets of posterior parameters include a false acceptance rate (FAR) biased threshold and a false rejection rate (FRR) biased threshold.
9. The audio processing circuit of claim 8 , wherein the wake-word threshold of the wake-word recognition engine is adjusted between the FAR biased threshold and the FRR biased threshold.
10. The audio processing circuit of claim 9 , wherein the wake-word threshold of the wake-word recognition engine is adjusted, based on the prediction data, to match current estimated usage conditions.
11. A method comprising:
generating always-on voice (AOV) prediction data for a current time via an AOV usage model of an audio processing circuit, wherein the AOV usage model is utilized as a function of time of day;
generating bias adjustment data corresponding to the current time of day, based on the AOV prediction data;
adjusting a wake-word threshold of a wake-word recognition engine of the audio processing circuit for the current time of day, based on the AOV prediction data; and
determining, via the wake-word recognition engine and based on audio data, when a wake-word is detected based on the generated bias adjustment data and the adjusted threshold.
12. The method of claim 11 , wherein the AOV usage model is generated based on historical data of AOV usage.
13. The method of claim 12 , wherein the historical data of AOV usage is based on times of wake-word detections.
14. The method of claim 13 , further comprising:
generating recognized query data based on the audio data after the wake-word is detected.
15. The method of claim 14 , wherein the historical data of AOV usage is further based on the recognized query data.
16. The method of claim 15 , wherein the wake-words are only considered in the historical data of AOV when followed by the recognized query data.
17. The method of claim 11 , wherein the wake-word recognition engine includes a model anchored by two sets of posterior parameters, only differing in threshold.
18. The method of claim 17 , wherein the two sets of posterior parameters include a false acceptance rate (FAR) biased threshold and a false rejection rate (FRR) biased threshold.
19. The method of claim 18 , wherein the wake-word threshold of the wake-word recognition engine is adjusted between the FAR biased threshold and the FRR biased threshold.
20. The method of claim 19 , wherein the wake-word threshold of the wake-word recognition engine is adjusted, based on the AOV prediction data, to match current estimated AOV usage conditions.