Reproduce “Contrastive Siamese Network for Semi-supervised Speech Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Soheil Khorram, Jaeyoung Kim, Anshuman Tripathi, Han Lu, Qian Zhang, Hasim Sak
A Contrastive Siamese (c-siam) Network from Google Inc. improves Automatic Speech Recognition performance by leveraging unlabeled audio data alongside limited supervised data, achieving up to 20% relative WER improvement over baselines, and mitigates the shortcut learning problem in transformer-based models through novel temporal augmentations.
Key insights
- Siamese networks can be successfully adapted to sequential speech data by addressing the "shortcut learning problem," where transformers exploit positional embeddings over acoustic content.
- Novel temporal augmentation techniques, such as uniform and non-uniform TempoAug, are crucial for forcing content-dependent learning and preventing reliance on positional information in transformer-based speech models.
- The stop-gradient mechanism from SimSiam can be effectively applied in speech representation learning to achieve collapse-free training in siamese networks without requiring negative pairs or momentum encoders.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.