wav2vec 2

A self-supervised framework that learns powerful speech representations from raw audio, drastically reducing the need for labeled training data.

Developed by Meta AI, wav2vec 2.0 processes raw audio through a convolutional encoder to construct latent representations, which are then masked and passed to a Transformer network. By solving a contrastive task over a quantized vocabulary of speech units, the model learns the underlying structure of spoken language without relying on manual transcriptions. This self-supervised approach allows developers to build highly accurate speech recognition systems: fine-tuning on just 10 minutes of labeled data yields a competitive 4.8% Word Error Rate (WER) on the LibriSpeech clean test set.

https://arxiv.org/abs/2006.11477