Techniques for language independent wake-up word detection

1. A method for training a user device, the method comprising:

receiving a first acoustic input of a user speaking a wake-up word in the target language;

providing a first acoustic feature derived from the first acoustic input to an acoustic model stored on the user device to obtain a first sequence of speech units corresponding to the wake-up word spoken by the user in the target language, the acoustic model trained on a corpus of training data in a source language different than the target language;

receiving a second acoustic input including the wake-up word in the target language;

providing a second acoustic feature derived from the second acoustic input to the acoustic model to obtain a second sequence of speech units corresponding to the wake-up word in the target language;

comparing the first sequence of speech units with the second sequence of speech units to recognize the wake-up word in the target language; and

adapting the acoustic model to the user using both the first and second sequence of speech units based on a comparison result.

2. The method of claim 1 , further comprising:

storing the first sequence of speech units as a reference characterizing the wake-up word.

3. The method of claim 1 , wherein the first acoustic input includes a first noise, the method further comprising:

adapting the acoustic model using the first noise.

4. The method of claim 3 , wherein the second acoustic input includes a second noise of a different type from the first noise.

5. The method of claim 1 , further comprising:

performing a speaker adaptation based on the first acoustic input using at least one of: a maximum a posteriori (MAP) adaptation, a maximum likelihood linear regression (MLLR) adaption, a linear input network (LIN) layer, a vocal tract length normalization (VTLN), a cluster adaptation, or an eigenvoice technique.

6. The method of claim 1 , wherein the second acoustic input further includes a command and is received while the user device is in a low power mode, the method further comprising:

performing a task associated with the command without exiting the low power mode.

7. A user device, comprising:

one or more microphones;

a computer readable medium; and

one or more processors programmed to:

responsive to receiving, via the one or more microphones, a first acoustic input of a user speaking a wake-up word in the target language, derive a first acoustic feature from the first acoustic input and provide the first acoustic feature to an acoustic model stored in the computer readable medium to obtain a first sequence of speech units corresponding to the wake-up word spoken by the user in the target language, the acoustic model trained on a corpus of training data in a source language different than the target language;

responsive to receiving, via the one or more microphones, a second acoustic input including the wake-up word in the target language while the user device is in a low power mode, derive a second acoustic feature from the second acoustic input and provide the second acoustic feature to the acoustic model to obtain a second sequence of speech units corresponding to the wake-up word in the target language,

compare the first sequence of speech units with the second sequence of speech units to recognize the wake-up word in the target language; and

adapt the acoustic model to the user using both the first and second sequence of speech units based on a comparison result.

8. The user device of claim 7 , wherein the one or more processors are further programmed to:

storing the first sequence of speech units in the computer readable medium as a reference characterizing the wake-up word.

9. The user device of claim 7 , wherein the first acoustic input includes a first noise, the one or more processors are further programmed to:

adapting the acoustic model using the first noise.

10. The user device of claim 9 , wherein the second acoustic input includes a second noise of a different type from the first noise.

11. The user device of claim 7 , wherein the one or more processors are further programmed to:

performing a speaker adaptation based on the first acoustic input using at least one of: a maximum a posteriori (MAP) adaptation, a maximum likelihood linear regression (MLLR) adaption, a linear input network (LIN) layer, a vocal tract length normalization (VTLN), a cluster adaptation, or an eigenvoice technique.

12. The user device of claim 7 , wherein the second acoustic input further includes a command, the one or more processors are further programmed to:

performing a task associated with the command without exiting the low power mode.

13. A method for a user device, comprising:

receiving an enrollment utterance of a user speaking a wake-up word in the target language;

deriving an enrollment acoustic feature from the enrollment utterance;

providing the enrollment acoustic feature to an acoustic model stored on the user device to obtain an enrollment sequence of speech units corresponding to the wake-up word spoken by the user in the target language, the acoustic model trained on a corpus of training data in a source language different than the target language;

receiving a command utterance including the wake-up word in the target language;

deriving a command acoustic feature from the commanding utterance;

providing the command acoustic feature to the acoustic model to obtain a command sequence of speech units corresponding to the wake-up word in the target language;

comparing the enrollment sequence of speech units with the command sequence of speech units to recognize the wake-up word in the target language; and

adapting the acoustic model to the user using both the enrollment and command sequence of speech units based on a comparison result.

14. The method of claim 13 , further comprising:

storing the enrollment sequence of speech units as a reference characterizing the wake-up word.

15. The method of claim 13 , wherein the enrollment utterance includes with a first noise, the method further comprising:

adapting the acoustic model using the first noise.

16. The method of claim 15 , wherein the command utterance includes a second noise of a different type from the first noise.

17. The method of claim 13 , wherein the command utterance further includes a command and is received while the user device is in a low power mode, the method further comprising:

performing a task associated with the command without exiting the low power mode.

18. The method of claim 13 , further comprising:

performing a speaker adaptation based on the enrollment acoustic input using at least one of: a maximum a posteriori (MAP) adaptation, or a maximum likelihood linear regression (MLLR) adaption.

19. The method of claim 13 , further comprising:

performing a speaker adaptation based on the enrollment acoustic input using at least one of: a linear input network (LIN) layer, or a vocal tract length normalization (VTLN).

20. The method of claim 13 , further comprising:

performing a speaker adaptation based on the enrollment acoustic input using at least one of: a cluster adaptation, or an eigenvoice technique.