Reproduce “Context-Aware Interleaved Batching for WhisperX” 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
Carlos Bain, Max Bain
University of Oxford researchers developed Context-Aware Interleaved Batching for WhisperX, a method that integrates continuous historical text context into batched automatic speech recognition. This approach yields an 8.4x speedup over sequential Whisper while reducing the Word Error Rate by 0.2 percentage points and improving proper noun transcription by 2.8 percentage points on the Earnings-21 dataset compared to standard WhisperX.
Key insights
- Autoregressive text conditioning can be successfully decoupled from internal timestamp decoding, preventing hallucination loops and error propagation common in standard Whisper.
- Structuring batched transcription by interleaving processing across parallel streams enables efficient throughput while preserving chronological textual context for each stream.
- Contextual information significantly improves the transcription of specific elements like proper nouns, technical terms, and punctuation across segment boundaries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.