Reproduce “Depth through Recurrence: Towards Ultra-Efficient On-Device ASR” 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
Chen Feng, Tianyi Xu, Ramchalam Kinattinkara Ramakrishnan, Yicheng Lin, Shaojie Zhuo, Zhaocong Yuan, Chenzheng Su, Xiaopeng Zhang
The authors compress speech-recognition encoders by grouping contiguous layers with similar representations, measured using CKA, and sharing one block across each group. They initialize shared blocks from teacher weights and train students with distillation. On STT-480ms, a two-block student uses 14.2M encoder parameters versus 108.9M for the teacher, with mean WER of 13.40% versus 11.05% across eight benchmarks, so it does not match teacher WER.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.