Temporal Taxation Compounds Under Post-Training Compression of Whisper Models
Temporal Taxation Compounds Under Post-Training Compression of Whisper Models
Whisper 模型在训练后压缩过程中的“时间税”加剧效应
Abstract: Automatic speech recognition models are audited for demographic fairness at full precision, yet the models that ship to production have been quantized, pruned, and distilled. We ask whether post-training weight compression, which alters model weights rather than the audio signal or its feature representation, redistributes error burden across demographic groups.
摘要： 自动语音识别模型通常在全精度下进行人口统计学公平性审计，然而实际部署到生产环境的模型往往经过了量化、剪枝和蒸馏处理。我们探讨了训练后的权重压缩（即改变模型权重而非音频信号或其特征表示）是否会重新分配不同人口群体间的错误负担。
Across the Whisper family on Fair-Speech, Common Voice 25, and AfriSpeech-200, 50% Wanda pruning of Whisper-large-v3 sharply widens the Black/AA-vs-Asian temporal-taxation differential on Fair-Speech: the absolute word-error-rate gap between the worst- and best-served groups more than doubles; at an assumed cost of five seconds of correction effort per transcription error this is a rise from 30 to 64 seconds of correction time per minute of speech.
在 Fair-Speech、Common Voice 25 和 AfriSpeech-200 数据集上对 Whisper 系列模型进行测试发现，对 Whisper-large-v3 进行 50% 的 Wanda 剪枝会显著扩大 Fair-Speech 数据集中黑人/非裔美国人与亚裔群体之间的“时间税”差异：表现最差群体与表现最好群体之间的绝对词错误率（WER）差距扩大了一倍以上；若假设每个转录错误需要 5 秒的校正成本，则每分钟语音的校正时间从 30 秒增加到了 64 秒。
This +111% relative increase is invariant to the assumed per-error cost, survives an audio-quality control, and is only partly mitigated by beam-search decoding, which still leaves an +86% increase. At edge model size, INT4 HQQ quantization compounds catastrophic transcript loops on West African accents by factors of five to seven.
这种 111% 的相对增长与假设的单位错误成本无关，在音频质量控制下依然存在，且仅能通过束搜索（beam-search）解码部分缓解，缓解后仍有 86% 的增长。在边缘模型尺寸下，INT4 HQQ 量化导致西非口音的灾难性转录循环增加了 5 到 7 倍。
Distillation, by contrast, narrows demographic gaps in 21 of 27 evaluated settings (teacher-student pair, precision, and dataset), with the exceptions concentrated on a single model pair. We cast the temporal-taxation construct of Choi and Choi (2025) as a quantitative metric, and show that single-snapshot fairness audits on full-precision models do not capture the deployment-time burden that compression places on already-marginalized speakers.
相比之下，蒸馏技术在 27 种评估设置（教师-学生对、精度和数据集）中的 21 种里缩小了人口统计学差距，仅在单一模型对上出现了例外。我们将 Choi 和 Choi (2025) 提出的“时间税”概念转化为定量指标，并证明对全精度模型进行的单次快照公平性审计，无法捕捉到压缩技术给本已边缘化的说话人带来的部署时负担。