研究者在对OpenAI的Whisper语音识别模型进行详细测试后发现，该模型对老年人的转录准确度实际上优于年轻人。[1]基于Common Voice 17英文数据集中的2,760个音频片段的分析显示，70多岁人群的词错率为4.67%，相比20多岁人群的6.53%降低了1.86个百分点。[1]
然而，研究同时揭示了端点检测环节存在的年龄偏差问题。[1]当采用700毫秒的固定沉默阈值进行语音断点判断时，年长者被系统过早中断的风险显著更高。[1]具体而言，60多岁人群的被打断率为19.7%，较20多岁的8.0%增加11.6个百分点；70多岁人群为16.6%，增加8.5个百分点；80多岁人群更是达到22.1%，相比20多岁的4.1%增长5.4倍。[1]这一现象的根本原因在于年长者的说话模式差异——他们平均每段话中包含2次停顿，总停顿时间约420-480毫秒，而20多岁人群仅有1次停顿，总时间240毫秒。[1]
研究进一步测试了改进方案的效果。[1]使用语义转折模型smart-turn v3后，60多岁和70多岁人群的打断率差距分别从+11.6个百分点和+8.5个百分点降至+5.9个百分点和+4.0个百分点，表明智能端点检测能显著缓解年龄相关的识别中断问题。[1]
Researchers testing OpenAI's Whisper speech recognition model on 2,760 audio clips from the Common Voice dataset discovered a counterintuitive finding: the system transcribes speakers in their 70s more accurately than those in their 20s [1]. The word error rate for speakers in their 70s reached 4.67%, compared to 6.53% for those in their 20s—a reduction of 1.86 percentage points [1]. The study controlled for accent and gender variations by analyzing 920 matched audio samples per age group from the Common Voice 17 English dataset [1].
However, the research revealed a significant challenge in how the model handles endpoint detection, particularly for older speakers [1]. When using a fixed silence threshold of 700 milliseconds to determine when speakers finish talking, older adults experienced substantially higher interruption rates [1]. Speakers in their 60s were interrupted 19.7% of the time compared to 8.0% for those in their 20s—an increase of 11.6 percentage points [1]. The disparity widened for even older participants: those in their 80s faced interruption rates of 22.1%, more than five times higher than the 4.1% rate for 20-year-olds [1].
The underlying cause stems from natural speech patterns in aging [1]. Older speakers averaged two pauses per utterance with a combined duration of 420 to 480 milliseconds, while younger speakers typically paused just once for around 240 milliseconds [1]. Implementing a semantic turn-taking model called smart-turn v3 partially mitigated this issue, reducing the interruption rate gap for speakers in their 60s from 11.6 percentage points to 5.9 percentage points, and for those in their 70s from 8.5 percentage points to 4.0 percentage points [1].