Streaming ASR quality is materially worse than offline ASR

Streaming ASR systems (whisper-streaming, WhisperLive, etc.) produce transcription quality that is noticeably worse than offline Whisper on the same audio. Word error rates are higher, especially at segment boundaries.

Streaming ASR must emit tokens with bounded latency as audio arrives, which requires chunked inference with causal masking and finite look-ahead. The model cannot see future audio context, so it makes decisions with incomplete information. The LocalAgreement algorithm and other streaming policies trade quality for latency.

Use offline ASR when latency is not critical. For streaming, tune the latency/quality tradeoff parameters (look-ahead window, chunk size) and accept the quality gap.

This is a fundamental tradeoff, not a bug. The Whisper-Streaming paper (arXiv 2307.14743) reports 3.3 seconds latency on unsegmented long-form speech but does not claim parity with offline Whisper. The quality gap is most pronounced at word boundaries and in noisy conditions.