Low-latency conversational large language models
A method includes receiving a transcription of an utterance, processing, using a first model, the transcription to generate a first text segment that represents an initial portion of a response to the utterance, processing, using a TTS system, the first text segment to generate a first synthesized speech representation, and providing, for audible output, the first synthesized speech representation. The method also includes providing, to a second model different from the first model, the transcription and the first text segment, the second model comprising an LLM configured to process the transcription and the first text segment to generate a second text segment that represents a remaining portion of the response to the utterance. The method further includes obtaining a second synthesized speech representation generated from the second text segment, and providing, for audible output by the user device, the second synthesized speech representation.
1 . A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:
receiving a transcription of an utterance spoken by a user of a user device;
processing, using a trained first model, the transcription to generate a first text segment that represents an initial portion of a response to the utterance;
processing, using a text-to-speech system, the first text segment to generate a first synthesized speech representation of the initial portion of the response to the utterance;
providing, for audible output by the user device, the first synthesized speech representation;
providing, to a trained second model different from the trained first model, the transcription and the first text segment, the trained second model comprising a large language model (LLM) configured to process the transcription and the first text segment to generate a second text segment that represents a remaining portion of the response to the utterance;
obtaining a second synthesized speech representation generated from the second text segment, the second synthesized speech representation representing the remaining portion of the response to the utterance; and
after the first synthesized speech representation of the initial portion of the response to the utterance is audibly output by the user device, providing, for audible output from the user device, the second synthesized speech representation,
wherein the first synthesized speech representation and the second synthesized speech representation audibly output by the user device are both response to the same utterance spoken by the user of the user device.
2 . The computer-implemented method of claim 1 , wherein the operations further comprise:
receiving audio data characterizing the utterance, the audio data captured by the user device; and
processing, using an automatic speech recognition system, the audio data to generate the transcription.
3 . The computer-implemented method of claim 1 , wherein:
the data processing hardware resides on the user device;
the trained first model executes on the data processing hardware; and
the trained second model executes on a remote computing system in communication with the data processing hardware.
4 . The computer-implemented method of claim 1 , wherein:
the data processing hardware resides on a remote computing system in communication with the user device; and
the trained first model and the second model both execute on the data processing hardware.
5 . The computer-implemented method of claim 1 , wherein the trained second model is trained on training transcriptions, each training transcription paired with a corresponding training initial response portion to condition the second model to learn how to generate a response to the training transcription that incorporates the corresponding training initial response portion.
6 . The computer-implemented method of claim 1 , wherein the trained first model is trained to generate the first text segment of one or more initial words in the response to the transcription such that the first synthesized speech representation generated from the one or more initial words includes a duration sufficient to mask a latency time period incurred while the trained second model processes the transcription and the first text segment to generate the second text segment.
7 . The computer-implemented method of claim 6 , wherein the one or more initial words represent at least one of a generic phrase, a filler phrase, or a prefix phrase.
8 . The computer-implemented method of claim 1 , wherein the trained first model comprises a classifier model configured to select, based on the transcription, the first text segment from a plurality of pre-determined text segments.
9 . The computer-implemented method of claim 1 , wherein:
the trained first model comprises a first LLM; and
the LLM of the trained second model comprises a second LLM having a greater number of parameters than the first LLM.
10 . The computer-implemented method of claim 1 , wherein the trained first model comprises an embedding model configured to project the transcription into an embedding space corresponding to a plurality of pre-determined first text segments.
11 . The computer-implemented method of claim 1 , wherein providing, for audible output from the user device, the second synthesized speech representation comprises discontinuing providing, for audible output by the user device, a remaining portion of the first synthesized speech representation.
12 . The computer-implemented method of claim 1 , wherein the first synthesized speech representation is provided for audible output while at least one of providing the transcription and the first text segment to the trained second model or obtaining the second synthesized speech representation.
13 . A system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware and storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:
receiving a transcription of an utterance spoken by a user of a user device;
processing, using a trained first model, the transcription to generate a first text segment that represents an initial portion of a response to the utterance;
processing, using a text-to-speech system, the first text segment to generate a first synthesized speech representation of the initial portion of the response to the utterance;
providing, for audible output by the user device, the first synthesized speech representation;
providing, to a trained second model different from the trained first model, the transcription and the first text segment, the trained second model comprising a large language model (LLM) configured to process the transcription and the first text segment to generate a second text segment that represents a remaining portion of the response to the utterance;
obtaining a second synthesized speech representation generated from the second text segment, the second synthesized speech representation representing the remaining portion of the response to the utterance; and
after the first synthesized speech representation of the initial portion of the response to the utterance is audibly output by the user device, providing, for audible output from the user device, the second synthesized speech representation,
wherein the first synthesized speech representation and the second synthesized speech representation audibly output by the user device are both response to the same utterance spoken by the user of the user device.
14 . The system of claim 13 , wherein the operations further comprise:
receiving audio data characterizing the utterance, the audio data captured by the user device; and
processing, using an automatic speech recognition system, the audio data to generate the transcription.
15 . The system of claim 13 , wherein:
the data processing hardware resides on the user device;
the trained first model executes on the data processing hardware; and
the trained second model executes on a remote computing system in communication with the data processing hardware.
16 . The system of claim 13 , wherein:
the data processing hardware resides on a remote computing system in communication with the user device; and
the trained first model and the second model both execute on the data processing hardware.
17 . The system of claim 13 , wherein the trained first model is trained to generate the first text segment of one or more initial words in the response to the transcription such that the first synthesized speech representation generated from the one or more initial words includes a duration sufficient to mask a latency time period incurred while the trained second model processes the transcription and the first text segment to generate the second text segment.
18 . The system of claim 13 , wherein the trained first model is trained to generate the first text segment of one or more initial words in the response to the transcription such that the first synthesized speech representation generated from the one or more initial words includes a duration sufficient to mask a latency time period incurred while the trained second model processes the transcription and the first text segment to generate the second text segment.
19 . The system of claim 18 , wherein the one or more initial words represent at least one of a generic phrase, a filler phrase, or a prefix phrase.
20 . The system of claim 13 , wherein the trained first model comprises a classifier model configured to select, based on the transcription, the first text segment from a plurality of pre-determined text segments.
21 . The system of claim 13 , wherein:
the trained first model comprises a first LLM; and
the LLM of the trained second model comprises a second LLM having a greater number of parameters than the first LLM.
22 . The system of claim 13 , wherein the trained first model comprises an embedding model configured to project the transcription into an embedding space corresponding to a plurality of pre-determined first text segments.
23 . The system of claim 13 , wherein providing, for audible output by the user device, the second synthesized speech representation comprises discontinuing providing, for audible output by the user device, a remaining portion of the first synthesized speech representation.
24 . The system of claim 13 , wherein the first synthesized speech representation is provided for audible output while at least one of providing the transcription and the first text segment to the trained second model or obtaining the second synthesized speech representation.