Context-Aware Interleaved Batching for WhisperX
The paper proposes keeping Whisper’s running text context while still batching segmented audio for speed.
The authors say WhisperX gains throughput by batching within audio, but that isolation can hurt punctuation and terminology. Their Context-Aware Interleaved Batching uses VAD-derived segment boundaries to preserve historical context across batched segments without triggering Whisper’s sequential slowdown and hallucination loops. On long-form audio benchmarks, they report lower WER and better proper noun transcription while maintaining high-throughput inference. ArXiv · AI/CL/LG's note
The authors say WhisperX gains throughput by batching within audio, but that isolation can hurt punctuation and terminology. Their Context-Aware Interleaved Batching uses VAD-derived segment boundaries to preserve historical context across batched segments without triggering Whisper’s sequential slowdown and hallucination loops. On long-form audio benchmarks, they report lower WER and better proper noun transcription while maintaining high-throughput inference. ArXiv · AI/CL/LG's note
score 4