Customizable speech recognition system

A computer-implemented method comprising:

• training an automatic speech recognition neural network using a generic dataset to output characters, wherein the training includes:
• receiving a training utterance related to the generic domain,
• generating a training hidden representations based on the training utterance,
• using the training hidden representations to generate a training context vector that indicates a training predicted character related to the training utterance,
• generating the training predicted character,
• determining training error in the automatic speech recognition neural network based on the training predicted character and
• initially training the automatic speech recognition neural network by updating the automatic speech recognition neural network based on the training error;
• training a language model neural network using the generic dataset;
• freezing one or more encoder layers of the automatic speech recognition neural network;
• adapting the automatic speech recognition neural network for a target dataset by maintaining weights for the one or more frozen encoder layers and updating weights of one or more unfrozen decoder layers based on the target dataset;
• adapting the language model neural network to the target dataset;
• and
• outputting a customized speech recognition neural network system comprised of the adapted automatic speech recognition neural network and the adapted language model neural network.

A computing system comprising:

• means for obtaining an utterance related to a target domain, the utterance represented as a dimension vector;
• means for generating, using an automatic speech recognition neural network, a context vector based on the dimension vector, the context vector indicating a predicted character related to the utterance, wherein the automatic speech recognition neural network is configured to generate character units for input utterances related to the target domain by maintaining parameters of a layer of the automatic speech recognition neural network associated with a generic domain during adaptation of the automatic speech recognition neural network, wherein one or more outputs from the automatic speech recognition neural network reach a threshold level of accuracy for the target domain;
• and
• means for generating, using the automatic speech recognition neural network, a character unit from the context vector that represents the predicted character.