posts_details 😃

Some posts not in my site

Low-resource ASR research

Speech representation: contrastive learning, generative learning and other methods

1. VQ-CPC code: Vector Quantization loss (reconstruction of Conv1d encoded input and rnn output) and CPC loss (cross-entropy loss of predictor from rnn output and negative samples with label 0, aiming to adjacement output and encoder output) discrete representation with CPC 🌊

Augmentation:

Transfer learning:

Multilingual ASR

Embedding multilingual information in discrete representaion like codebook.
Paper: UNSUPERVISED PRETRAINING TRANSFERS WELL ACROSS LANGUAGES:CPC for speech representation across languages pre-training