Augmenting automated speech recognition neural networks with scalable vocabularies
Techniques for augmenting automated speech recognition neural networks with scalable vocabularies are described. A cluster is selected from a plurality of clusters of similar sounding words based on a score, the score representing a similarity between an embedding of the cluster and an audio embedding of an utterance generated with an automated speech recognition encoder neural network. A bias factor is calculated based on a similarity between an embedding of a word in the selected cluster and the audio embedding. The audio embedding of the utterance is biased by the bias factor.
1 . A computer-implemented method comprising:
clustering words in a vocabulary into a plurality of clusters of similar sounding words using k-means clustering, wherein a number of clusters of the plurality of clusters is based on a maximum cluster size;
calculating, with a first attention neural network, a similarity score for each cluster based on a dot product of an embedding of each cluster with an audio embedding of an utterance generated with an automated speech recognition encoder neural network;
selecting a cluster from the plurality of clusters having the highest similarity score;
calculating, by an embedding encoder neural network, an embedding of a word in the selected cluster based at least in part on at least one of an audio sample of a pronunciation of the word or a phonetic representation of the word;
calculating, with a second attention neural network, a bias factor based on a similarity between the embedding of the word and the audio embedding of the utterance;
biasing the audio embedding of the utterance by the bias factor to generate a biased audio embedding of the utterance; and
generating, by a machine learning service based at least in part on the biased audio embedding of the utterance, a text prediction of the utterance.
2 . The computer-implemented method of claim 1 , wherein generating the text prediction of the utterance comprises processing the biased audio embedding of the utterance using a linear layer or a fully-connected layer, and a softmax operation, followed by a beam search decoder to collapse the biased audio embedding of the utterance into the text prediction of the utterance.
3 . The computer-implemented method of claim 1 , wherein all words in the plurality of clusters of similar sounding words are part of a vocabulary identified in a request received by the machine learning service.
4 . A computer-implemented method comprising:
calculating, with a first attention neural network, a score for a cluster based on a dot product of an embedding of the cluster and an audio embedding of an utterance generated with an automated speech recognition encoder neural network;
selecting the cluster from a plurality of clusters of similar sounding words based on the score, wherein the score represents a similarity between the embedding of the cluster and the audio embedding of the utterance;
calculating, with a second attention neural network, a similarity between an embedding of a word in the cluster and the audio embedding of the utterance based on a dot product of the embedding of the word and the audio embedding of the utterance, wherein the first attention neural network and the second attention neural network have the same weight parameters;
calculating a bias factor based on the similarity between the embedding of the word in the cluster and the audio embedding of the utterance;
biasing the audio embedding of the utterance by the bias factor to generate a biased audio embedding of the utterance; and
generating, by a machine learning service based at least in part on the biased audio embedding of the utterance, a text prediction of the utterance.
5 . The computer-implemented method of claim 4 , wherein the automated speech recognition encoder neural network was trained with training data, and wherein the first attention neural network was trained using a subset of the training data, the subset of the training data based on samples in the training data including words having a frequency in the training data below a threshold.
6 . The computer-implemented method of claim 4 , further comprising calculating, with an embedding encoder neural network, the embedding of the word in the selected cluster that positions the embedding of the word in a multi-dimensional embedding space near other similar sounding words.
7 . The computer-implemented method of claim 6 , wherein the selected cluster includes a plurality of clustered words and further comprising:
calculating, with the embedding encoder neural network, an embedding of each word in the plurality of clustered words,
wherein the embedding of the cluster is based on an average of each embedding of each word in the plurality of clustered words.
8 . The computer-implemented method of claim 4 , wherein generating the text prediction of the utterance comprises processing the biased audio embedding of the utterance using a linear layer or a fully-connected layer, and a softmax operation, followed by a beam search decoder to collapse the biased audio embedding of the utterance into the text prediction of the utterance.
9 . The computer-implemented method of claim 4 , wherein all words in the plurality of clusters of similar sounding words are part of a vocabulary identified in a request received by the machine learning service.
10 . The computer-implemented method of claim 9 , further comprising clustering the words in the vocabulary into the plurality of clusters of similar sounding words using k-means clustering, wherein a maximum number of words in each cluster is limited by a maximum cluster size.
11 . The computer-implemented method of claim 4 , wherein biasing the audio embedding of the utterance by the bias factor comprises at least one of adding the bias factor to the audio embedding or scaling the audio embedding by the bias factor.
12 . A system comprising:
a first one or more electronic devices to implement a storage service in a multi-tenant provider network to store word cluster data including a plurality of clusters of similar sounding words and, for each cluster, an embedding of the cluster; and
a second one or more electronic devices to implement a model hosting service in the multi-tenant provider network, the model hosting service including instructions that upon execution cause the model hosting service to:
retrieve the word cluster data from the storage service;
calculate, with a first attention neural network, a score for a cluster based on a dot product of an embedding of the cluster and an audio embedding of an utterance generated with an automated speech recognition encoder neural network;
select the cluster from the plurality of clusters of similar sounding words based on the score, wherein the score represents a similarity between the embedding of the cluster and the audio embedding of the utterance;
calculate, with a second attention neural network, a similarity between an embedding of a word in the cluster and the audio embedding of the utterance based on a dot product of the embedding of the word and the audio embedding of the utterance, wherein the first attention neural network and the second attention neural network have the same weight parameters;
calculate a bias factor based on the similarity between the embedding of the word in the cluster and the audio embedding of the utterance;
bias the audio embedding of the utterance by the bias factor to generate a biased audio embedding of the utterance; and
generate, based at least in part on the biased audio embedding of the utterance, a text prediction of the utterance.
13 . The system of claim 12 , wherein the model hosting service includes further instructions that upon execution cause the model hosting service to calculate, with an embedding encoder neural network, the embedding of the word in the selected cluster that positions the embedding of the word in a multi-dimensional embedding space near other similar sounding words.
14 . The system of claim 13 , wherein the selected cluster includes a plurality of clustered words; and
wherein the model hosting service includes further instructions that upon execution cause the model hosting service to calculate, with the embedding encoder neural network, an embedding of each word in the plurality of clustered words,
wherein the embedding of the cluster is based on an average of each embedding of each word in the plurality of clustered words.
15 . The system of claim 12 , wherein the automated speech recognition encoder neural network was trained with training data, and wherein the first attention neural network was trained using a subset of the training data, the subset of the training data based on samples in the training data including words having a frequency in the training data below a threshold.
16 . The system of claim 12 , wherein generating the text prediction of the utterance comprises processing the biased audio embedding of the utterance using a linear layer or a fully-connected layer, and a softmax operation, followed by a beam search decoder to collapse the biased audio embedding of the utterance into the text prediction of the utterance.
17 . The system of claim 12 , wherein all words in the plurality of clusters of similar sounding words are part of a vocabulary identified in a request received by the machine learning service.
18 . The system of claim 17 , wherein the model hosting service includes further instructions that upon execution cause the model hosting service to cluster the words in the vocabulary into the plurality of clusters of similar sounding words using k-means clustering, wherein a maximum number of words in each cluster is limited by a maximum cluster size.
19 . The system of claim 12 , wherein biasing the audio embedding of the utterance by the bias factor comprises at least one of adding the bias factor to the audio embedding or scaling the audio embedding by the bias factor.