Abstract: A hypothesis stitcher for speech recognition of long-form audio provides superior performance, such as higher accuracy and reduced computational cost. An example disclosed operation includes: segmenting the audio stream into a plurality of audio segments; identifying a plurality of speakers within each of the plurality of audio segments; performing automatic speech recognition (ASR) on each of the plurality of audio segments to generate a plurality of short-segment hypotheses; merging at least a portion of the short-segment hypotheses into a first merged hypothesis set; inserting stitching symbols into the first merged hypothesis set, the stitching symbols including a window change (WC) symbol; and consolidating, with a network-based hypothesis stitcher, the first merged hypothesis set into a first consolidated hypothesis.

Abstract: Embodiments are associated with a speaker-independent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-independent attention-based encoder-decoder model associated with a first output distribution, and a speaker-dependent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-dependent attention-based encoder-decoder model associated with a second output distribution. The second attention-based encoder-decoder model is trained to classify output tokens based on input speech frames of a target speaker and simultaneously trained to maintain a similarity between the first output distribution and the second output distribution.

Abstract: Solutions for on-device streaming inverse text normalization (ITN) include: receiving a stream of tokens, each token representing an element of human speech; tagging, by a tagger that can work in a streaming manner (e.g., a neural network), the stream of tokens with one or more tags of a plurality of tags to produce a tagged stream of tokens, each tag of the plurality of tags representing a different normalization category of a plurality of normalization categories; based on at least a first tag representing a first normalization category, converting, by a first language converter of a plurality of category-specific natural language converters (e.g., weighted finite state transducers, WFSTs), at least one token of the tagged stream of tokens, from a first lexical language form, to a first natural language form; and based on at least the first natural language form, outputting a natural language representation of the stream of tokens.

Abstract: The disclosure herein describes using a transcript generation model for generating a transcript from a multi-speaker audio stream. Audio data including overlapping speech of a plurality of speakers is obtained and a set of frame embeddings are generated from audio data frames of the obtained audio data using an audio data encoder. A set of words and channel change (CC) symbols are generated from the set of frame embeddings using a transcript generation model. The CC symbols are included between pairs of adjacent words that are spoken by different people at the same time. The set of words and CC symbols are transformed into a plurality of transcript lines, wherein words of the set of words are sorted into transcript lines based on the CC symbols, and a multi-speaker transcript is generated based on the plurality of transcript lines. The inclusion of CC symbols by the model enables efficient, accurate multi-speaker transcription.

Abstract: A computing system including one or more processors configured to receive an audio input. The one or more processors may generate a text transcription of the audio input at a sequence-to-sequence speech recognition model, which may assign a respective plurality of external-model text tokens to a plurality of frames included in the audio input. Each external-model text token may have an external-model alignment within the audio input. Based on the audio input, the one or more processors may generate a plurality of hidden states. Based on the plurality of hidden states, the one or more processors may generate a plurality of output text tokens. Each output text token may have a corresponding output alignment within the audio input. For each output text token, a latency between the output alignment and the external-model alignment may be below a predetermined latency threshold. The one or more processors may output the text transcription.

Abstract: A hypothesis stitcher for speech recognition of long-form audio provides superior performance, such as higher accuracy and reduced computational cost. An example disclosed operation includes: segmenting the audio stream into a plurality of audio segments; identifying a plurality of speakers within each of the plurality of audio segments; performing automatic speech recognition (ASR) on each of the plurality of audio segments to generate a plurality of short-segment hypotheses; merging at least a portion of the short-segment hypotheses into a first merged hypothesis set; inserting stitching symbols into the first merged hypothesis set, the stitching symbols including a window change (WC) symbol; and consolidating, with a network-based hypothesis stitcher, the first merged hypothesis set into a first consolidated hypothesis.

Abstract: A hypothesis stitcher for speech recognition of long-form audio provides superior performance, such as higher accuracy and reduced computational cost. An example disclosed operation includes: segmenting the audio stream into a plurality of audio segments; identifying a plurality of speakers within each of the plurality of audio segments; performing automatic speech recognition (ASR) on each of the plurality of audio segments to generate a plurality of short-segment hypotheses; merging at least a portion of the short-segment hypotheses into a first merged hypothesis set; inserting stitching symbols into the first merged hypothesis set, the stitching symbols including a window change (WC) symbol; and consolidating, with a network-based hypothesis stitcher, the first merged hypothesis set into a first consolidated hypothesis.

Abstract: A computing system including one or more processors configured to receive an audio input. The one or more processors may generate a text transcription of the audio input at a sequence-to-sequence speech recognition model, which may assign a respective plurality of external-model text tokens to a plurality of frames included in the audio input. Each external-model text token may have an external-model alignment within the audio input. Based on the audio input, the one or more processors may generate a plurality of hidden states. Based on the plurality of hidden states, the one or more processors may generate a plurality of output text tokens. Each output text token may have a corresponding output alignment within the audio input. For each output text token, a latency between the output alignment and the external-model alignment may be below a predetermined latency threshold. The one or more processors may output the text transcription.

Abstract: A computer device is provided that includes one or more processors configured to receive an end-to-end (E2E) model that has been trained for automatic speech recognition with training data from a source-domain, and receive an external language model that has been trained with training data from a target-domain. The one or more processors are configured to perform an inference of the probability of an output token sequence given a sequence of input speech features. Performing the inference includes computing an E2E model score, computing an external language model score, and computing an estimated internal language model score for the E2E model. The estimated internal language model score is computed by removing a contribution of an intrinsic acoustic model. The processor is further configured to compute an integrated score based at least on E2E model score, the external language model score, and the estimated internal language model score.

Abstract: A hypothesis stitcher for speech recognition of long-form audio provides superior performance, such as higher accuracy and reduced computational cost. An example disclosed operation includes: segmenting the audio stream into a plurality of audio segments; identifying a plurality of speakers within each of the plurality of audio segments; performing automatic speech recognition (ASR) on each of the plurality of audio segments to generate a plurality of short-segment hypotheses; merging at least a portion of the short-segment hypotheses into a first merged hypothesis set; inserting stitching symbols into the first merged hypothesis set, the stitching symbols including a window change (WC) symbol; and consolidating, with a network-based hypothesis stitcher, the first merged hypothesis set into a first consolidated hypothesis.

Abstract: A computer device is provided that includes one or more processors configured to receive an end-to-end (E2E) model that has been trained for automatic speech recognition with training data from a source-domain, and receive an external language model that has been trained with training data from a target-domain. The one or more processors are configured to perform an inference of the probability of an output token sequence given a sequence of input speech features. Performing the inference includes computing an E2E model score, computing an external language model score, and computing an estimated internal language model score for the E2E model. The estimated internal language model score is computed by removing a contribution of an intrinsic acoustic model. The processor is further configured to compute an integrated score based at least on E2E model score, the external language model score, and the estimated internal language model score.

Abstract: Embodiments are associated with a speaker-independent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-independent attention-based encoder-decoder model associated with a first output distribution, and a speaker-dependent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-dependent attention-based encoder-decoder model associated with a second output distribution. The second attention-based encoder-decoder model is trained to classify output tokens based on input speech frames of a target speaker and simultaneously trained to maintain a similarity between the first output distribution and the second output distribution.

Abstract: The disclosure herein describes training a global model based on a plurality of data sets. The global model is applied to each data set of the plurality of data sets and a plurality of gradients is generated based on that application. At least one gradient quality metric is determined for each gradient of the plurality of gradients. Based on the determined gradient quality metrics of the plurality of gradients, a plurality of weight factors is calculated. The plurality of gradients is transformed into a plurality of weighted gradients based on the calculated plurality of weight factors and a global gradient is generated based on the plurality of weighted gradients. The global model is updated based on the global gradient, wherein the updated global model, when applied to a data set, performs a task based on the data set and provides model output based on performing the task.

Abstract: Embodiments are associated with a speaker-independent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-independent attention-based encoder-decoder model associated with a first output distribution, a speaker-dependent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-dependent attention-based encoder-decoder model associated with a second output distribution, training of the second attention-based encoder-decoder model to classify output tokens based on input speech frames of a target speaker and simultaneously training the speaker-dependent attention-based encoder-decoder model to maintain a similarity between the first output distribution and the second output distribution, and performing automatic speech recognition on speech frames of the target speaker using the trained speaker-dependent attention-based encoder-decoder model.

Abstract: A computing system including one or more processors configured to receive an audio input. The one or more processors may generate a text transcription of the audio input at a sequence-to-sequence speech recognition model, which may assign a respective plurality of external-model text tokens to a plurality of frames included in the audio input. Each external-model text token may have an external-model alignment within the audio input. Based on the audio input, the one or more processors may generate a plurality of hidden states. Based on the plurality of hidden states, the one or more processors may generate a plurality of output text tokens. Each output text token may have a corresponding output alignment within the audio input. For each output text token, a latency between the output alignment and the external-model alignment may be below a predetermined latency threshold. The one or more processors may output the text transcription.

Abstract: Described herein are systems and methods for a general, scalable, end-to-end framework that uses a generative adversarial network (GAN) objective to enable robust speech recognition. Encoders trained with the proposed approach enjoy improved invariance by learning to map noisy audio to the same embedding space as that of clean audio. Embodiments of a Wasserstein GAN framework increase the robustness of seq-to-seq models in a scalable, end-to-end fashion. In one or more embodiments, an encoder component is treated as the generator of GAN and is trained to produce indistinguishable embeddings between labeled and unlabeled audio samples. This new robust training approach can learn to induce robustness without alignment or complicated inference pipeline and even where augmentation of audio data is not possible.

Abstract: Embodiments are associated with a speaker-independent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-independent attention-based encoder-decoder model associated with a first output distribution, a speaker-dependent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-dependent attention-based encoder-decoder model associated with a second output distribution, training of the second attention-based encoder-decoder model to classify output tokens based on input speech frames of a target speaker and simultaneously training the speaker-dependent attention-based encoder-decoder model to maintain a similarity between the first output distribution and the second output distribution, and performing automatic speech recognition on speech frames of the target speaker using the trained speaker-dependent attention-based encoder-decoder model.

Abstract: Described herein are systems and methods to identify and address sources of bias in an end-to-end speech model. In one or more embodiments, the end-to-end model may be a recurrent neural network with two 2D-convolutional input layers, followed by multiple bidirectional recurrent layers and one fully connected layer before a softmax layer. In one or more embodiments, the network is trained end-to-end using the CTC loss function to directly predict sequences of characters from log spectrograms of audio. With optimized recurrent layers and training together with alignment information, some unwanted bias induced by using purely forward only recurrences may be removed in a deployed model.

Abstract: Described herein are systems and methods for a general, scalable, end-to-end framework that uses a generative adversarial network (GAN) objective to enable robust speech recognition. Encoders trained with the proposed approach enjoy improved invariance by learning to map noisy audio to the same embedding space as that of clean audio. Embodiments of a Wasserstein GAN framework increase the robustness of seq-to-seq models in a scalable, end-to-end fashion. In one or more embodiments, an encoder component is treated as the generator of GAN and is trained to produce indistinguishable embeddings between labeled and unlabeled audio samples. This new robust training approach can learn to induce robustness without alignment or complicated inference pipeline and even where augmentation of audio data is not possible.

Abstract: Described herein are systems and methods to identify and address sources of bias in an end-to-end speech model. In one or more embodiments, the end-to-end model may be a recurrent neural network with two 2D-convolutional input layers, followed by multiple bidirectional recurrent layers and one fully connected layer before a softmax layer. In one or more embodiments, the network is trained end-to-end using the CTC loss function to directly predict sequences of characters from log spectrograms of audio. With optimized recurrent layers and training together with alignment information, some unwanted bias induced by using purely forward only recurrences may be removed in a deployed model.