Reproduce “A Better and Faster End-to-End Model for Streaming 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
Bo Li, Anmol Gulati, Jiahui Yu, Tara N. Sainath, Chung-Cheng Chiu, Arun Narayanan, Shuo-Yiin Chang, Ruoming Pang and 7 more
Google LLC researchers developed an end-to-end streaming ASR model that integrates a Conformer encoder, FastEmit, E2E Prefetching, and Cascaded Encoders to improve both recognition quality and latency for applications like Voice Search. The proposed two-pass model achieved a 5.4% Word Error Rate (WER) and a median prefetch latency of 0ms, outperforming previous streaming E2E models and conventional hybrid systems across key metrics.
Key insights
- Conformer architectures can be successfully adapted for multi-domain, streaming ASR by implementing specific modifications like Group Normalization and local self-attention, leading to robust quality gains.
- Sequence-level emission regularization like FastEmit provides an effective, alignment-free, and computationally efficient method to significantly reduce partial latency in RNN-T models with minimal impact on WER.
- A unified two-pass decoding approach with Cascaded Encoders can effectively recover quality by incorporating non-causal information from the second pass while remaining compatible with streaming ASR's prefetching mechanisms, yielding a superior overall quality-latency profile.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.