Speech similarity embeddings with sensitivity to recording conditions via two-stage augmentations

A computer implemented method comprising:

• for a plurality of training examples from a training dataset, each training example comprising a pair of audio snippets:
• processing, by a machine-learning embedding model, a representation of a first audio snippet in the pair of audio snippets to generate an encoded representation of the first audio snippet;
• processing, by the machine-learning embedding model, a representation of a second audio snippet in the pair of audio snippets to generate an encoded representation of the second audio snippet;
• and
• comparing the encoded representation of the first audio snippet to the encoded representation of the second audio snippet using a contrastive loss function;
• and
• updating parameters of the machine-learning embedding model based at least in part on the comparisons of the encoded representations of the first audio snippets to the encoded representations of the second audio snippets,
• wherein the training dataset comprises a plurality training examples sampled from an augmented dataset, the augmented dataset comprising a plurality of augmented audio snippets, wherein an augmented audio snippet is generated by applying one or more augmentations from a second set of augmentations to an audio snippet taken from an underlying audio sample that has had one or more augmentations from a first set of augmentations applied, wherein the first set of augmentations are stronger than the second set of augmentations.