Generation of training examples for training automatic speech recognizers

A computer system comprising:

• a processor and a memory configured to provide computer program instructions to the processor to execute a function of components of an example generator including:
• a training dataset receiving a training dataset including a plurality of original audio signals;
• an imperceptible space component for constructing an imperceptible space for an original audio signal using a psychoacoustic model;
• an adversarial example checking component for searching for candidate audio examples in the imperceptible space by checking candidate audio examples with an automatic speech recognizer to determine which candidate audio example meets a difference level to be an adversarial audio example compared to the original audio signal;
• an adversarial training example component for generating adversarial training examples from the original audio signal in the training dataset for training an automatic speech recognizer, the adversarial training examples being generated based, at least in part, on the imperceptible space constructed for the original audio signal and finding adversarial audio examples in the imperceptible space;
• and
• an example generator component for inputting the adversarial audio example to an adversarial trainer for training the automatic speech recognizer, the adversarial audio example being both adversarial and imperceptible.

A computer program product comprising:

• one or more computer readable storage media and program instructions stored on the one or more computer readable storage media, the program instructions comprising:
• program instructions to receive a training dataset including a plurality of original audio signals;
• and
• program instructions to generate adversarial training examples from an original audio signal in the training dataset for training an automatic speech recognizer, the adversarial training examples being generated based, at least in part, on an imperceptible space constructed for an original audio signal of the original audio signals and finding adversarial audio examples in the imperceptible space, wherein the program instructions to generate training examples for training an automatic speech recognize comprise:
• program instructions to construct the imperceptible space for the original audio signal using a psychoacoustic model,
• program instructions to search for candidate audio examples in the imperceptible space;
• and
• program instructions to, for a set of candidate audio examples identified in the imperceptible space, determine, via the automatic speech recognizer, candidate audio example meets a difference level to be an adversarial audio example compared to the original audio signal;
• and
• program instructions to input the adversarial audio example to an adversarial trainer for training the automatic speech recognizer, the adversarial audio example being both adversarial and imperceptible.