USING LARGE LANGUAGE MODEL(S) IN GENERATING AUTOMATED ASSISTANT RESPONSE(S)
As part of a dialog session between a user and an automated assistant, implementations can receive a stream of audio data that captures a spoken utterance including an assistant query, determine, based on processing the stream of audio data, a set of assistant outputs that are each predicted to be responsive to the assistant query, process, using large language model (LLM) output(s), the assistant outputs and context of the dialog session to generate a set of modified assistant outputs, and cause given modified assistant output, from among the set of modified assistant outputs, to be provided for presentation to the user in response to the spoken utterance. In some implementations, the LLM output(s) can be generated in an offline manner for subsequent use in an online manner. In additional or alternative implementations, the LLM output(s) can be generated in an online manner when the spoken utterance is received.
1 . A method implemented by one or more processors, the method comprising:
as part of a dialog session between a user of a client device and an automated assistant implemented by the client device:
receiving a stream of audio data that captures a spoken utterance of the user, the stream of audio data being generated by one or more microphones of the client device, and the spoken utterance including an assistant query;
determining, based on processing the stream of audio data, a set of assistant outputs, each assistant output in the set of assistant outputs being responsive to the assistant query included in the spoken utterance;
processing, using a large language model (LLM), the set of assistant outputs and context of the dialog session to generate a set of modified assistant outputs that each reflect a first personality from among a plurality of disparate personalities; and
causing a given modified assistant output, from among the set of modified assistant outputs, to be provided for presentation to the user.
2 . The method of claim 1 , wherein processing the set of assistant outputs and the context of the dialog session to generate the set of modified assistant outputs that each reflect the first personality and using the LLM causes the LLM to adapt the set of assistant output to a first vocabulary, that is associated with the first personality, to generate the set of modified assistant outputs.
3 . The method of claim 2 , wherein the first personality is disparate from a second personality, and wherein the first vocabulary, that is associated with the first personality and that is used to generate the set of modified assistant outputs, is disparate from a second vocabulary, that is associated with the second personality.
4 . The method of claim 1 , wherein causing the given modified assistant output to be provided for presentation to the user comprises:
generating, using a first set of prosodic properties that is associated with the first personality, synthesized speech audio data that captures the given modified assistant output; and
causing the synthesized speech audio data that captures the given modified assistant output to be audibly rendered via one or more speakers of the client device.
5 . The method of claim 4 , wherein the first personality is disparate from a second personality, and wherein the first set of prosodic properties, that is associated with the first personality and that is used to generate the synthesized speech audio data, is disparate from a second set of prosodic properties, that is associated with the second personality.
6 . The method of claim 1 , further comprising:
selecting the given modified assistant output, from among the set of modified assistant outputs, based on a probability distribution over a sequence of words or phrases that is generated using the LLM.
7 . The method of claim 1 , wherein the LLM is a first LLM, from among a plurality of disparate LLMs, that is specific to the first personality.
8 . The method of claim 1 , wherein the first personality is defined by the user in settings of a software application.
9 . The method of claim 1 , further comprising:
as part of a subsequent dialog session between the user and the automated assistant implemented by the client device:
receiving a subsequent stream of audio data that captures a subsequent spoken utterance of the user, the subsequent stream of audio data being generated by one or more of the microphones of the client device, and the subsequent spoken utterance including a subsequent assistant query;
determining, based on processing the subsequent stream of audio data, a subsequent set of assistant outputs, each assistant output in the subsequent set of assistant outputs being responsive to the subsequent assistant query included in the subsequent spoken utterance;
processing, using the LLM or an additional LLM, the subsequent set of assistant outputs and subsequent context of the subsequent dialog session to generate a subsequent set of modified assistant outputs that each reflect a second personality that differs from the first personality; and
causing a given subsequent modified assistant output, from among the subsequent set of modified assistant outputs, to be provided for presentation to the user.
10 . The method of claim 9 , wherein the subsequent set of assistant outputs and the subsequent context of the subsequent dialog session are processed using the additional LLM to generate the subsequent set of modified assistant outputs that each reflect the second personality that differs from the first personality, wherein the LLM is associated with the first personality, and wherein the additional LLM is associated with the first personality.
11 . A system comprising:
at least one processor; and
memory storing instructions that, when executed, cause the at least one processor to be operable to:
as part of a dialog session between a user of a client device and an automated assistant implemented by the client device:
receive a stream of audio data that captures a spoken utterance of the user, the stream of audio data being generated by one or more microphones of the client device, and the spoken utterance including an assistant query;
determine, based on processing the stream of audio data, a set of assistant outputs, each assistant output in the set of assistant outputs being responsive to the assistant query included in the spoken utterance;
process, using a large language model (LLM), the set of assistant outputs and context of the dialog session to generate a set of modified assistant outputs that each reflect a first personality from among a plurality of disparate personalities; and
cause a given modified assistant output, from among the set of modified assistant outputs, to be provided for presentation to the user.
12 . The system of claim 11 , wherein processing the set of assistant outputs and the context of the dialog session to generate the set of modified assistant outputs that each reflect the first personality and using the LLM causes the LLM to adapt the set of assistant output to a first vocabulary, that is associated with the first personality, to generate the set of modified assistant outputs.
13 . The system of claim 12 , wherein the first personality is disparate from a second personality, and wherein the first vocabulary, that is associated with the first personality and that is used to generate the set of modified assistant outputs, is disparate from a second vocabulary, that is associated with the second personality.
14 . The system of claim 11 , wherein the instructions to cause the given modified assistant output to be provided for presentation to the user comprise instructions to:
generate, using a first set of prosodic properties that is associated with the first personality, synthesized speech audio data that captures the given modified assistant output; and
cause the synthesized speech audio data that captures the given modified assistant output to be audibly rendered via one or more speakers of the client device.
15 . The system of claim 14 , wherein the first personality is disparate from a second personality, and wherein the first set of prosodic properties, that is associated with the first personality and that is used to generate the synthesized speech audio data, is disparate from a second set of prosodic properties, that is associated with the second personality.
16 . The system of claim 11 , wherein the at least one processor is further operable to:
select the given modified assistant output, from among the set of modified assistant outputs, based on a probability distribution over a sequence of words or phrases that is generated using the LLM.
17 . The method of claim 1 , wherein the LLM is a first LLM, from among a plurality of disparate LLMs, that is specific to the first personality.
18 . The system of claim 11 , wherein the first personality is defined by the user in settings of a software application.
19 . The system of claim 1 , wherein the at least one processor is further operable to:
as part of a subsequent dialog session between the user and the automated assistant implemented by the client device:
receive a subsequent stream of audio data that captures a subsequent spoken utterance of the user, the subsequent stream of audio data being generated by one or more of the microphones of the client device, and the subsequent spoken utterance including a subsequent assistant query;
determine, based on processing the subsequent stream of audio data, a subsequent set of assistant outputs, each assistant output in the subsequent set of assistant outputs being responsive to the subsequent assistant query included in the subsequent spoken utterance;
process, using the LLM or an additional LLM, the subsequent set of assistant outputs and subsequent context of the subsequent dialog session to generate a subsequent set of modified assistant outputs that each reflect a second personality that differs from the first personality; and
cause a given subsequent modified assistant output, from among the subsequent set of modified assistant outputs, to be provided for presentation to the user.
20 . A non-transitory computer-readable storage medium storing instructions that, when executed by at least one processor, cause the at least one processor to be operable to perform operations, the operations comprising:
as part of a dialog session between a user of a client device and an automated assistant implemented by the client device:
receiving a stream of audio data that captures a spoken utterance of the user, the stream of audio data being generated by one or more microphones of the client device, and the spoken utterance including an assistant query;
determining, based on processing the stream of audio data, a set of assistant outputs, each assistant output in the set of assistant outputs being responsive to the assistant query included in the spoken utterance;
processing, using a large language model (LLM), the set of assistant outputs and context of the dialog session to generate a set of modified assistant outputs that each reflect a first personality from among a plurality of disparate personalities; and
causing a given modified assistant output, from among the set of modified assistant outputs, to be provided for presentation to the user.