Predictive pre-recording of audio for voice input

1. A system comprising:

one or more computers; and

a computer-readable medium coupled to the one or more computers having instructions stored thereon which, when executed by the one or more computers, cause the one or more computers to perform operations comprising:

training a probabilistic model using data that reflects whether, given past combinations of (i) state data that references a state of a mobile device and (ii) sensor data that is sensed by one or more sensors of the mobile device, a user of the mobile device did or did not actually initiate a voice input;

obtaining, as input data, particular state data that references a particular state of a mobile device, and particular sensor data that is sensed by one or more sensors of the mobile device;

inferring, based on applying the probabilistic model to the input data, that the user of the mobile device is likely to initiate voice input; and

invoking voice input functionality of the mobile device in response to inferring that the user is likely to initiate voice input.

2. The system of claim 1 , wherein invoking voice input functionality of the mobile device further comprises commencing a background audio recording.

3. The system of claim 1 , wherein the particular state data comprises data that references whether a display of the mobile device is turned on or turned off, data that references whether the mobile device is operating in a locked mode or in an unlocked mode, data that references one or more applications that are executing on the mobile device, data that references whether a voice search application is executing on the mobile device, data that references whether a field selected by the user is enabled for voice text entry, or data that references whether the user is operating an input mechanism of the mobile device.

4. The system of claim 1 , wherein the particular state data comprises data that references a current state or a recent state of the mobile device.

5. The system of claim 1 , wherein the particular sensor data comprises data that is sensed by a keypad sensor of the mobile device, data that is sensed by a position determination sensor of the mobile device, data that is sensed by an accelerometer of the mobile device, data that is sensed by a magnetometer of the mobile device, data that is sensed by a light sensor of the mobile device, data that is sensed by a proximity sensor of the mobile device, data that is sensed by a capacitive sensor of the mobile device, or data that is sensed by a touch sensor of the mobile device.

6. The system of claim 1 , wherein:

establishing, as input data, particular sensor data further comprises establishing, as the input data, data that is sensed by a first sensor of the mobile device, and

invoking voice input functionality of the mobile device further comprises activating a second sensor of the mobile device in response to inferring that the user is likely to initiate voice input.

7. The system of claim 6 , wherein the operations further comprise:

establishing, as additional input data, state data, data that is sensed by the first sensor of the mobile device, and data that is sensed by the second sensor of the mobile device;

applying an additional probabilistic model to the additional input data;

inferring, based on applying the additional probabilistic model to the additional input data, that the user remains likely to initiate voice input; and

commencing a background recording in response to inferring that the user remains likely to initiate voice input.

8. The system of claim 1 , wherein:

applying the probabilistic model to the input data further comprises determining a score associated with the input data; and

inferring that the user is likely to initiate voice input further comprises determining that the score associated with the input data satisfies a predefined threshold.

9. The system of claim 1 , wherein the operations further comprise:

inferring, based on applying the probabilistic model to the input data, that it is unclear whether the user is likely to initiate voice input;

instructing the mobile device to collect additional state data or additional sensor data in response to inferring that it is unclear whether the user is likely to initiate voice input; and

applying the probabilistic model to the input data and the additional state data or the additional sensor data,

wherein inferring that the user is likely to initiate voice input is further based on applying the probabilistic model to the input data and the additional state data or the additional sensor data.

10. The system of claim 1 , wherein the operations further comprise:

receiving historical input data collected by other mobile devices; and

training the probabilistic model using the historical input data associated with the mobile device and the historical input data collected by the other mobile devices.

11. The system of claim 1 , wherein inferring that the user is likely to initiate voice input further comprises inferring, by the mobile device, that the user is likely to initiate voice input.

12. The system of claim 1 , wherein inferring that the user is likely to initiate voice input further comprises inferring, by a server, that the user is likely to initiate voice input.

13. The system of claim 1 , wherein the operations further comprise:

communicating an audio signal and duration data between the mobile device and the server, wherein the duration data indicates a duration of a background audio recording encoded in the audio signal that is recorded before the user begins speaking.

14. A computer-implemented method comprising:

training a probabilistic model using data that reflects whether, given past combinations of (i) state data that references a state of a mobile device and (ii) sensor data that is sensed by one or more sensors of the mobile device, a user of the mobile device did or did not actually initiate a voice input;

obtaining, as input data, particular state data that references a particular state of a mobile device, and particular sensor data that is sensed by one or more sensors of the mobile device;

inferring, based on applying the probabilistic model to the input data, that the user of the mobile device is likely to initiate voice input; and

invoking voice input functionality of the mobile device in response to inferring that the user is likely to initiate voice input.

15. The method of claim 14 , wherein:

establishing, as input data, particular sensor data further comprises establishing, as the input data, data that is sensed by a first sensor of the mobile device, and

invoking voice input functionality of the mobile device further comprises activating a second sensor of the mobile device in response to inferring that the user is likely to initiate voice input.

16. The method of claim 15 , further comprising:

establishing, as additional input data, state data, data that is sensed by the first sensor of the mobile device, and data that is sensed by the second sensor of the mobile device;

applying an additional probabilistic model to the additional input data;

inferring, based on applying the additional probabilistic model to the additional input data, that the user remains likely to initiate voice input; and

commencing a background recording in response to inferring that the user remains likely to initiate voice input.

17. The method of claim 14 , wherein:

applying the probabilistic model to the input data further comprises determining a score associated with the input data; and

inferring that the user is likely to initiate voice input further comprises determining that the score associated with the input data satisfies a predefined threshold.

18. A non-transitory computer storage medium encoded with a computer program, the program comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

training a probabilistic model using data that reflects whether, given past combinations of (i) state data that references a state of a mobile device and (ii) sensor data that is sensed by one or more sensors of the mobile device, a user of the mobile device did or did not actually initiate a voice input;

obtaining, as input data, particular state data that references a particular state of a mobile device, and particular sensor data that is sensed by one or more sensors of the mobile device;

inferring, based on applying the probabilistic model to the input data, that the user of the mobile device is likely to initiate voice input; and

invoking voice input functionality of the mobile device in response to inferring that the user is likely to initiate voice input.