RT Neil Zeghidour：基准测试的首要目标是促进模型改进。我们加入了Coval基准测试，以领先的延迟率参赛，然后取得了很大的进步……
RT Neil Zeghidour: The first goal of a benchmark is to foster model improvement. We joined the coval benchmark with leading latency and then got much ...
Matt TurckAI2026-09-10
RT Neil Zeghidour 基准测试的第一个目标是促进模型改进。我们加入了Coval基准测试，以领先的延迟率参与，然后排名大幅下降。这是因为延迟现在是根据可听到的第一个音频而不是第一个音频数据包来衡量的。这让我们意识到我们的第一个音频数据包是完全静音的，现在已修复！感谢@covaldev让我们的模型变得更好。Coval：大多数TTS延迟基准测试在任何音频到达时（无论是否静音）就会停止计时。我们改变了我们的方法，将静音也计算在内。更新发布的那天，@gradiumai报告说延迟从172毫秒增加到430毫秒。没有模型更改或基础设施更改。他们的音频到达了
原文
RT Neil ZeghidourThe first goal of a benchmark is to foster model improvement. We joined the coval benchmark with leading latency and then got much worse rankings. This was because latency was now measured based on first hearable audio rather than first audio packet. That made us realize our first audio packets were completely silent, now fixed! Thanks @covaldev for making our models better.coval: Most TTS latency benchmarks stop the clock the moment any audio arrives, silent or not. We changed ours to count that silence. The day the update shipped, @gradiumai reported latency moved from 172 ms to 430 ms. No model change or infrastructure change. Their audio was arriving