Language models using domain-specific model components

1. A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:

obtaining a plurality of training language examples for training a language model to recognize speech in a particular domain represented by a combination of multiple different aspects of non-linguistic context, wherein:

each training language example occurs in one or more of the multiple different aspects of non-linguistic context representing the particular domain; and

the language model comprises:

a baseline model component; and

multiple domain-specific model components each corresponding to a respective different aspect of non-linguistic context from the multiple different aspects of non-linguistic context representing the particular domain;

training, using the plurality of training language examples, the language model by updating corresponding weights of the multiple domain-specific model components;

obtaining an utterance comprising a non-linguistic context; and

determining a transcription of the utterance using the language model by:

determining a score of a candidate transcription of the utterance using the baseline model component;

adjusting the score of the candidate transcription using at least one domain-specific model component of the multiple domain-specific model components of the language model, wherein the at least one domain-specific model component is selected based on the non-linguistic context; and

determining the transcription for the utterance based on the adjusted score.

2. The method of claim 1 , wherein training the language model by updating the corresponding weights of the multiple domain-specific model components comprises training the language model by updating the corresponding weights of the multiple domain-specific model components without updating corresponding weights of the baseline model components.

3. The method of claim 1 , wherein the baseline model component is domain independent.

4. The method of claim 1 , wherein the baseline model component comprises corresponding weights for a respective set of features.

5. The method of claim 1 , wherein each aspect of non-linguistic context from the multiple different aspects of non-linguistic context corresponds to at least one of a location, a time condition, a user characteristic, a device characteristic, or a device status.

6. The method of claim 1 , wherein the baseline model component comprises a log-linear model comprising corresponding weights for a corresponding set of features.

7. The method of claim 6 , wherein the corresponding weights of the baseline model component are for features that represent occurrence of n-grams independent of non-linguistic context.

8. The method of claim 1 , wherein each of the multiple domain-specific model components are log-linear models that each comprise the corresponding weights for a corresponding set of features.

9. A system comprising:

data processing hardware; and

memory hardware in communication with the data processing hardware and storing instructions that when executed by the data processing hardware cause the data processing hardware to perform operations comprising:

obtaining a plurality of training language examples for training a language model to recognize speech in a particular domain represented by a combination of multiple different aspects of non-linguistic context, wherein:

each training language example occurs in one or more of the multiple different aspects of non-linguistic context representing the particular domain; and

the language model comprises:

a baseline model component; and

multiple domain-specific model components each corresponding to a respective different aspect of non-linguistic context from the multiple different aspects of non-linguistic context representing the particular domain;

training, using the plurality of training language examples, the language model by updating corresponding weights of the multiple domain-specific model components;

obtaining an utterance comprising a non-linguistic context; and

determining a transcription of the utterance using the language model by:

determining a score of a candidate transcription of the utterance using the baseline model component;

adjusting the score of the candidate transcription using at least one domain-specific model component of the multiple domain-specific model components of the language model, wherein the at least one domain-specific model component is selected based on the non-linguistic context; and

determining the transcription for the utterance based on the adjusted score.

10. The system of claim 9 , wherein training the language model by updating the corresponding weights of the multiple domain-specific model components comprises training the language model by updating the corresponding weights of the multiple domain-specific model components without updating corresponding weights of the baseline model components.

11. The system of claim 9 , wherein the baseline model component is domain independent.

12. The system of claim 9 , wherein the baseline model component comprises corresponding weights for a respective set of features.

13. The system of claim 9 , wherein each aspect of non-linguistic context from the multiple different aspects of non-linguistic context corresponds to at least one of a location, a time condition, a user characteristic, a device characteristic, or a device status.

14. The system of claim 9 , wherein the baseline model component comprises a log-linear model comprising corresponding weights for a corresponding set of features.

15. The system of claim 14 , wherein the corresponding weights of the baseline model component are for features that represent occurrence of n-grams independent of non-linguistic context.

16. The system of claim 9 , wherein each of the multiple domain-specific model components are log-linear models that each comprise the corresponding weights for a corresponding set of features.