Abstract: Implementations relate to reducing latency in generating and/or rendering a given stream of natural language (NL) based output generated using a large language model (LLM). Processor(s) of a system can: receive NL based input associated with a client device, generate the stream of NL based output utilizing the LLM that is responsive to the NL based input and that is for a given dialog context of an ongoing dialog, and cause the stream of NL based output to be rendered at the client device. Notably, the processor(s) can employ attribute classifier(s) and a multi-objective scorer to implement a blockwise controlled decoding technique in generating the stream of NL based output utilizing the LLM. By implementing the blockwise controlled decoding technique in generating the stream of NL based output utilizing the LLM, the processor(s) can reduce latency in generating and/or of the stream of NL based output generated utilizing the LLM.

Abstract: Implementations disclose selecting, in response to receiving a request and from among multiple candidate generative models (e.g., multiple candidate large language models (LLMs)) with differing computational efficiencies, a particular generative model to utilize in generating a response to the request. Those implementations reduce latency and/or conserve computational resource(s) through selection, for various requests, of a more computationally efficient generative model for utilization in lieu of a less computationally efficient generative model. Further, those implementations seek to achieve such benefits, through utilization of more computationally efficient generative models, while also still selectively utilizing less computationally efficient generative models for certain requests to mitigate occurrences of a generated response being inaccurate and/or under-specified.

Abstract: Implementations relate to handling visual content across a multi-turn dialog. A user input that includes natural language content and visual content is received during the dialog. If the visual content is being received for the first time in the dialog, the visual content is processed to generate a corresponding tokenized representation of the visual content. The corresponding tokenized representation can be cached in a database in association with the dialog, or in association with a user account of a user of the user query. If the visual content is subsequently referenced in the dialog, the corresponding tokenized representation of the visual content is retrieved from the database. The corresponding tokenized representation of the visual content, corresponding tokenized representations of natural language content, and optionally other metadata can be processed, using a generative model, to generate a response responsive to the user input.

Abstract: A method includes receiving training data that includes a set of unspoken textual utterances. For each respective unspoken textual utterance, the method includes, tokenizing the respective textual utterance into a sequence of sub-word units, generating a first higher order textual feature representation for a corresponding sub-word unit tokenized from the respective unspoken textual utterance, receiving the first higher order textual feature representation generated by a text encoder, and generating a first probability distribution over possible text units. The method also includes training an encoder based on the first probability distribution over possible text units generated by a first-pass decoder for each respective unspoken textual utterance in the set of unspoken textual utterances.

Abstract: A method includes processing, using a speech recognizer, a first portion of audio data to generate a first lattice, and generating a first partial transcription for an utterance based on the first lattice. The method includes processing, using the recognizer, a second portion of the data to generate, based on the first lattice, a second lattice representing a plurality of partial speech recognition hypotheses for the utterance and a plurality of corresponding speech recognition scores. For each particular partial speech recognition hypothesis, the method includes generating a corresponding re-ranked score based on the corresponding speech recognition score and whether the particular partial speech recognition hypothesis shares a prefix with the first partial transcription.

Abstract: Implementations relate to reducing latency in generating and/or rendering natural language (NL) output generated using a large language model (LLM). Processor(s) of a system can: receive NL based input associated with a client device, and generate the NL based output utilizing the LLM. The NL based output can be a stream of NL based output in that it includes a plurality of segments, and is generated on a segment-by-segment basis. In some implementations, a first segment of the stream of NL based output is selected for inclusion in the stream of NL based output as a second segment (and any subsequent segment) is being generated to reduce latency in evaluating the NL based output as a whole prior to rendering thereof. In some versions of those implementations, the first segment is rendered as the second segment (and any subsequent segment) is being generated to further reduce latency in rendering thereof.

Abstract: A method includes receiving a sequence of acoustic frames as input to a multilingual automated speech recognition (ASR) model configured to recognize speech in a plurality of different supported languages and generating, by an audio encoder of the multilingual ASR, a higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The method also includes generating, by a language identification (LID) predictor of the multilingual ASR, a language prediction representation for a corresponding higher order feature representation. The method also includes generating, by a decoder of the multilingual ASR, a probability distribution over possible speech recognition results based on the corresponding higher order feature representation, a sequence of non-blank symbols, and a corresponding language prediction representation. The decoder includes monolingual output layer having a plurality of output nodes each sharing a plurality of language-specific wordpiece models.

Abstract: Two-pass automatic speech recognition (ASR) models can be used to perform streaming on-device ASR to generate a text representation of an utterance captured in audio data. Various implementations include a first-pass portion of the ASR model used to generate streaming candidate recognition(s) of an utterance captured in audio data. For example, the first-pass portion can include a recurrent neural network transformer (RNN-T) decoder. Various implementations include a second-pass portion of the ASR model used to revise the streaming candidate recognition(s) of the utterance and generate a text representation of the utterance. For example, the second-pass portion can include a listen attend spell (LAS) decoder. Various implementations include a shared encoder shared between the RNN-T decoder and the LAS decoder.

Abstract: Implementations relate to handling visual content across a multi-turn dialog. A user input that includes natural language content and visual content is received during the dialog. If the visual content is being received for the first time in the dialog, the visual content is processed to generate a corresponding tokenized representation of the visual content. The corresponding tokenized representation can be cached in a database in association with the dialog, or in association with a user account of a user of the user query. If the visual content is subsequently referenced in the dialog, the corresponding tokenized representation of the visual content is retrieved from the database. The corresponding tokenized representation of the visual content, corresponding tokenized representations of natural language content, and optionally other metadata can be processed, using a generative model, to generate a response responsive to the user input.

Abstract: A method includes obtaining a set of training samples, wherein each training sample includes a corresponding sequence of speech segments corresponding to a training utterance and a corresponding sequence of ground-truth transcriptions for the sequence of speech segments, and wherein each ground-truth transcription includes a start time and an end time of a corresponding speech segment. For each training sample in the set of training samples, the method includes processing, using a speech recognition model, the corresponding sequence of speech segments to obtain one or more speech recognition hypotheses for the training utterance; and, for each speech recognition hypothesis obtained for the training utterance, identifying a respective number of word errors relative to the corresponding sequence of ground-truth transcriptions.

Abstract: Implementations relate to dialog management of a large language model (LLM) utilized in generating natural language (NL) output during an ongoing dialog. Processor(s) of a system can: receive NL based input as part of the ongoing dialog, generate NL based output utilizing the LLM, and cause the NL based output to be rendered. Further, the processor(s) can receive subsequent NL based input as part of the ongoing dialog. In some implementations, the processor(s) can determine whether to modify a corresponding dialog context in generating subsequent NL based output, and modify the corresponding dialog context accordingly. For example, the processor(s) can restrict the corresponding dialog context, or supplant the corresponding dialog context with a corresponding curated dialog context. In additional or alternative implementations, the processor(s) can modify a corresponding NL based output threshold utilized in generating the subsequent NL based response to ensure the resulting NL based output is desirable.

Abstract: A method includes receiving a training example that includes audio data representing a spoken utterance and a ground truth transcription. For each word in the spoken utterance, the method also includes inserting a placeholder symbol before the respective word identifying a respective ground truth alignment for a beginning and an end of the respective word, determining a beginning word piece and an ending word piece, and generating a first constrained alignment for the beginning word piece and a second constrained alignment for the ending word piece. The first constrained alignment is aligned with the ground truth alignment for the beginning of the respective word and the second constrained alignment is aligned with the ground truth alignment for the ending of the respective word. The method also includes constraining an attention head of a second pass decoder by applying the first and second constrained alignments.

Abstract: Two-pass automatic speech recognition (ASR) models can be used to perform streaming on-device ASR to generate a text representation of an utterance captured in audio data. Various implementations include a first-pass portion of the ASR model used to generate streaming candidate recognition(s) of an utterance captured in audio data. For example, the first-pass portion can include a recurrent neural network transformer (RNN-T) decoder. Various implementations include a second-pass portion of the ASR model used to revise the streaming candidate recognition(s) of the utterance and generate a text representation of the utterance. For example, the second-pass portion can include a listen attend spell (LAS) decoder. Various implementations include a shared encoder shared between the RNN-T decoder and the LAS decoder.

Abstract: A method of text-only and semi-supervised training for deliberation includes receiving training data including unspoken textual utterances that are each not paired with any corresponding spoken utterance of non-synthetic speech, and training a deliberation model that includes a text encoder and a deliberation decoder on the unspoken textual utterances. The method also includes receiving, at the trained deliberation model, first-pass hypotheses and non-causal acoustic embeddings. The first-pass hypotheses is generated by a recurrent neural network-transducer (RNN-T) decoder for the non-causal acoustic embeddings encoded by a non-causal encoder. The method also includes encoding, using the text encoder, the first-pass hypotheses generated by the RNN-T decoder, and generating, using the deliberation decoder attending to both the first-pass hypotheses and the non-causal acoustic embeddings, second-pass hypotheses.

Abstract: An automated speech recognition (ASR) model includes a first encoder, a first encoder, a second encoder, and a second decoder. The first encoder receives, as input, a sequence of acoustic frames, and generates, at each of a plurality of output steps, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The first decoder receives, as input, the first higher order feature representation generated by the first encoder, and generates a first probability distribution over possible speech recognition hypotheses. The second encoder receives, as input, the first higher order feature representation generated by the first encoder, and generates a second higher order feature representation for a corresponding first higher order feature frame. The second decoder receives, as input, the second higher order feature representation generated by the second encoder, and generates a second probability distribution over possible speech recognition hypotheses.

Abstract: A computer-implemented method includes receiving audio data that corresponds to an utterance spoken by a user and captured by a user device. The method also includes processing the audio data to determine a candidate transcription that includes a sequence of tokens for the spoken utterance. Tor each token in the sequence of tokens, the method includes determining a token embedding for corresponding token, determining a n-gram token embedding for a previous sequence of n-gram tokens, and concatenating the token embedding and the n-gram token embedding to generate a concatenated output for the corresponding token. The method also includes rescoring the candidate transcription for the spoken utterance by processing the concatenated output generated for each corresponding token in the sequence of tokens.

Abstract: Implementations relate to dialog management of a large language model (LLM) utilized in generating natural language (NL) output during an ongoing dialog. Processor(s) of a system can: receive NL based input as part of the ongoing dialog, generate NL based output utilizing the LLM, and cause the NL based output to be rendered. Further, the processor(s) can receive subsequent NL based input as part of the ongoing dialog. In some implementations, the processor(s) can determine whether to modify a corresponding dialog context in generating subsequent NL based output, and modify the corresponding dialog context accordingly. For example, the processor(s) can restrict the corresponding dialog context, or supplant the corresponding dialog context with a corresponding curated dialog context. In additional or alternative implementations, the processor(s) can modify a corresponding NL based output threshold utilized in generating the subsequent NL based response to ensure the resulting NL based output is desirable.

Abstract: A method includes receiving, as input to a speech recognition model, audio data corresponding to a spoken utterance. The method also includes performing, using the speech recognition model, speech recognition on the audio data by, at each of a plurality of time steps, encoding, using an audio encoder, the audio data corresponding to the spoken utterance into a corresponding audio encoding, and decoding, using a speech recognition joint network, the corresponding audio encoding into a probability distribution over possible output labels. At each of the plurality of time steps, the method also includes determining, using an intended query (IQ) joint network configured to receive a label history representation associated with a sequence of non-blank symbols output by a final softmax layer, an intended query decision indicating whether or not the spoken utterance includes a query intended for a digital assistant.

Abstract: A method includes receiving a training example that includes audio data representing a spoken utterance and a ground truth transcription. For each word in the spoken utterance, the method also includes inserting a placeholder symbol before the respective word identifying a respective ground truth alignment for a beginning and an end of the respective word, determining a beginning word piece and an ending word piece, and generating a first constrained alignment for the beginning word piece and a second constrained alignment for the ending word piece. The first constrained alignment is aligned with the ground truth alignment for the beginning of the respective word and the second constrained alignment is aligned with the ground truth alignment for the ending of the respective word. The method also includes constraining an attention head of a second pass decoder by applying the first and second constrained alignments.

Abstract: A method includes generating, using an audio encoder, a higher-order feature representation for each acoustic frame in a sequence of acoustic frames; generating, using a decoder, based on the higher-order feature representation, a plurality of speech recognition hypotheses, each hypotheses corresponding to a candidate transcription of an utterance and having an associated first likelihood score; generating, using an external language model, for each speech recognition hypothesis, a second likelihood score; determining, using a learnable fusion module, for each speech recognition hypothesis, a set of fusion weights based on the higher-order feature representation and the speech recognition hypothesis; and generating, using the learnable fusion module, for each speech recognition hypothesis, a third likelihood score based on the first likelihood score, the second likelihood score, and the set of fusion weights, the audio encoder and decoder trained using minimum additive error rate training in the presence of t