Reproduce “StreamAlign: Streaming Text-Aligned Speech Tokenization” 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
Kang-wook Kim, Jinyoung Park, Jinsoo Kim, Sehun Lee, Sang Hoon Woo, Gunhee Kim
A new framework, STREAMALIGN, introduces streaming, text-aligned speech tokenization that reduces latency to 270 ms for real-time applications and preserves fine-grained acoustic details by aligning speech to LLM subwords. It enables spoken language models to achieve competitive performance in both acoustic and semantic tasks, outperforming existing baselines in speech reconstruction and continuation.
Key insights
- Word-guided character alignment effectively leverages the accuracy of word-level ASR while achieving the fine-grained control needed for character and subword alignment, mitigating vocabulary mismatch.
- LLM-subword-level aggregation, as opposed to word-level, is critical for preserving acoustic granularity and improving speech reconstruction quality.
- A lightweight word boundary classifier significantly reduces end-to-end latency for streaming applications without compromising recognition or reconstruction quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.