智能体
x.com ·
@aiDotEngineer
AI Engineer 语音专场：生产级语音 Agent 的延迟、准确率与错误率经验
AI Engineer 汇总语音专场要点：Tolan 响应延迟从 2 秒漂移到 2.5 秒即导致产品指标全面下滑；Plivo 把口述电话号码当作带校验的字段而非转写，准确率从约 30% 提升到 90% 以上且无需微调；Hamming AI 监控 1 万个生产语音 Agent，真实错误率接近 10%，对抗测试约每 5 个击破 1 个；EliseAI 让后台 Agent 执行慢工具调用并把结果注入上下文；实时语音模型听到“mm hmm”会中断并道歉，而真实通话中约五分之一时间双方同时说话。
AI推荐理由 汇总 AI Engineer 语音专场中的多个一线工程经验：Tolan 延迟从 2 秒漂移到 2.5 秒即拖垮产品指标；Plivo 将口述电话号码当作带校验的字段而非转写，准确率从约 30% 提升到 90% 以上且无需微调；Hamming AI 监控 1 万个生产语音 Agent，真实错误率接近 10%，对抗测试约每 5 个 Agent 击破 1 个；EliseAI 用后台 Agent 执行慢工具调用；实时语音模型对“mm hmm”会中断道歉。内容具体、可操作，对语音 Agent 开发者价值高，但为会议要点转述，缺少完整上下文与原始数据。
78
浏览 566
点赞 2
评论 0
转发 0
正文
These moments resonated:
- Tolan's response latency drifted from two seconds to two and a half. That half second tanked essentially every product metric, and users wrote in to say their companion had gotten slow.
- Plivo treats a spoken phone number as a typed field with a validator, not a transcript. Accuracy went from about 30 percent to the mid nineties with no fine tuning.
- Hamming AI monitors 10,000 production voice agents. The real error rate is nearer 10 percent, and adversarial testing breaks roughly one agent in five.
- EliseAI has a background agent make the slow tool call and drop the result into context, so the main model believes it made the call itself.
- Say mm hmm to a realtime voice model and it stops dead to apologize. On a real phone call, up to a fifth of the time both people are talking at once.
Voice playlist:
[youtube.com/watch?v=LZuWZR…](https://www.youtube.com/watch?v=LZuWZRze3MU&list=PLcfpQ4tk2k0VdE7NSMKkNqc2qUH_lhm8K)