SPRING-INX-DATA2VEC-AQC-MARATHI

About Model

Data2vec-aqc is a self-supervised learning (SSL) model for speech representation learning, specifically designed to improve Automatic Speech Recognition (ASR) in low-resource settings. It is built using the Fairseq toolkit and extends the original data2vec framework by introducing three key modules: a quantizer (similar to wav2vec 2.0), a clustering module (from ccc-wav2vec 2.0), and a cross-contrastive loss mechanism. The model uses a Transformer-based architecture, consistent with data2vec, and operates in a teacher-student training setup. The student network processes randomly augmented versions of audio samples, while the teacher network (an exponentially moving average of the student) provides target representations. The student learns to predict the teacher’s contextualised latent representations, enabling robust feature learning.