Dynamic adaptation of language understanding systems to acoustic environments

1. A method for dynamic adaptation of a language understanding system to acoustic environments, the method comprising:

detecting, by executing instructions with a processor, a wake-on-voice key-phrase in or prior to an audio stream;

triggering, by executing instructions with the processor, a query recognition mode in response to the detecting of the wake-on-voice key-phrase;

performing, by executing instructions with the processor, automatic speech recognition on the audio stream to recognize a voice query;

applying, by executing instructions with the processor, a static language understanding classifier to the voice query, the static language understanding classifier employing a static semantic model;

applying, by executing instructions with the processor, a dynamic language understanding classifier to the voice query, the dynamic language understanding classifier employing a dynamic semantic model;

determining, by executing instructions with the processor, an intent associated with the voice query based on a linear interpolation of a first estimate of intent from the static language understanding classifier and a second estimate of intent from the dynamic language understanding classifier, the linear interpolation to combine the first estimate of intent and the second estimate of intent;

triggering, by executing instructions with the processor; a switch from the query recognition mode to an adaptation mode in response to determining the intent;

performing, by executing instructions with the processor, automatic speech recognition on the audio stream to generate hypotheses for audio events associated with utterances and background noise; and

updating, by executing instructions with the processor, the dynamic semantic model based on the hypotheses.

2. The method of claim 1 , wherein the generated hypotheses include one or more of an acoustic confidence level, a pitch level, or prosodic information.

3. The method of claim 1 , wherein the updating of the dynamic semantic model further includes performing a stochastic gradient descent optimization.

4. The method of claim 1 , further including updating the dynamic semantic model based on the voice query.

5. The method of claim 1 , wherein determining the intent associated with the recognized voice query is based on a class-constrained interpolation of the first estimate of intent from the static language understanding classifier and the second estimate of intent from the dynamic language understanding classifier.

6. A system comprising:

a wake-on-voice circuit to:

detect a wake-on-voice key-phrase in or prior to an audio stream,

trigger a query recognition mode in response to the detection of the wake-on-voice key-phrase, and

trigger a switch from the query recognition mode to an adaptation mode in response to determining an intent associated with a voice query;

an automatic speech recognition circuit to:

perform speech recognition on the audio stream during the query recognition mode,

recognize the voice query,

perform speech recognition on the audio stream during the adaptation mode, and

generate hypotheses for audio events associated with utterances and background noise;

a static language understanding classifier circuit to:

generate a first estimate of intent based on the voice query, and

employ a static semantic model;

a dynamic language understanding classifier circuit to:

generate a second estimate of intent based on the recognized in-domain query, and

employ a dynamic semantic model;

an interpolation circuit to determine an intent associated with the voice query based on a linear interpolation of the first estimate of intent from the static language understanding classifier circuit and the second estimate of intent from the dynamic language understanding classifier circuit, the linear interpolation including a combination of the first estimate of intent and the second estimate of intent; and

an adaptation circuit to update the dynamic semantic model based on the generated hypotheses.

7. The system of claim 6 , wherein the generated hypotheses include one or more of an acoustic confidence level, a pitch level, or prosodic information.

8. The system of claim 6 , wherein the adaptation circuit is to perform a stochastic gradient descent optimization to update the dynamic semantic model.

9. The system of claim 6 , wherein the adaptation circuit is to update the dynamic semantic model based on the voice query.

10. At least one non-transitory computer readable medium comprising instructions that, when executed, cause one or more processors to at least:

detect a wake-on-voice key-phrase in or prior to an audio stream;

trigger a query recognition mode in response to the detection of the wake-on-voice key-phrase;

perform automatic speech recognition on the audio stream to recognize a voice query;

apply a static language understanding classifier to the voice query, the static language understanding classifier to employ a static semantic model;

apply a dynamic language understanding classifier to the voice query, the dynamic language understanding classifier to employ a dynamic semantic model;

determine an intent associated with the voice query based on a linear interpolation of a first estimate of intent from the static language understanding classifier and a second estimate of intent from the dynamic language understanding classifier, the linear interpolation to combine the first estimate of intent and the second estimate of intent;

trigger a switch from the query recognition mode to an adaptation mode in response to determining the intent;

perform automatic speech recognition on the audio stream during the adaptation mode, to generate hypotheses for audio events associated with a rejection class that includes audio associated with out-of-domain utterances; and

update the dynamic semantic model based on the hypotheses.

11. The at least one non-transitory computer readable medium of claim 10 , wherein the generated hypotheses include one or more of an acoustic confidence level, a pitch level, or prosodic information.

12. The at least one non-transitory computer readable medium of claim 10 , wherein the updating of the dynamic semantic model includes performing a stochastic gradient descent optimization based on the voice query.

13. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions, when executed, cause the one or more processors to determine the intent associated with the voice query based on a class-constrained interpolation of the first estimate of intent from the static language understanding classifier and the second estimate of intent from the dynamic language understanding classifier.

14. The method of claim 1 , further including utilizing an influence weight in the linear interpolation to control an effect of the static language understanding classifier and the dynamic language understanding classifier on the intent.

15. The method of claim 5 , further including limiting the updating of the dynamic semantic model based on outlier cases including at least one of high-noise environments or semantic classes.

16. The system of claim 6 , wherein the interpolation circuit is to utilize an influence weight in the linear interpolation to control an effect of the static language understanding classifier and the dynamic language understanding classifier on the intent.

17. The system of claim 6 , wherein the interpolation circuit is to determine the intent associated with the voice query is based on a class-constrained interpolation of the first estimate of intent from the static language understanding classifier and the second estimate of intent from the dynamic language understanding classifier.

18. The system of claim 17 , wherein the adaptation circuit is to limit updating the dynamic semantic model to outlier cases including at least one of high-noise environments or semantic classes.

19. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions, when executed, cause the one or more processors to utilize an influence weight in the linear interpolation to control an effect of the static language understanding classifier and the dynamic language understanding classifier on the intent.

20. The at least one non-transitory computer readable medium of claim 13 , wherein the instructions, when executed, cause the one or more processors to limit updating of the dynamic semantic model to outlier cases including at least one of high-noise environments or semantic classes.

21. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions, when executed, cause the one or more processors to update the dynamic semantic model based on the voice query.

22. The at least one non-transitory computer readable medium of claim 10 , wherein the instructions, when executed, cause the one or more processors to use a first learning rate with the dynamic semantic model for a first environment and a second learning rate with the dynamic semantic model for a second environment, the first environment changing faster than the second environment, the first learning rate higher than the second learning rate.