跳转至

文章背景与核心概要

语音智能体在交互过程中常常难以精准判断用户何时结束发言，传统方法多依赖固定的静音超时机制，这往往会导致过早打断或响应迟钝。尽管针对多种语言的语义轮次结束（End-of-Turn, EOT）检测器已有所发展，但“TamilEOT”填补了南印度语种在该领域的空白，是首个针对泰米尔语的开源语义轮次结束检测资源。

该研究由 Santhoshkumar V 完成并于 2026 年 9 月发布。核心贡献包括包含 18,485 个标注轮次边界的 TamilEOT 数据集（从 116 场真实的泰米尔语电话对话中提取），以及基于 Smart Turn v3 微调的两个纯音频检测器。实验表明，相较于零样本基线，这些模型在准确率（从 70.30% 提升至 83.71% 和 86.13%）和 ROC-AUC（从 0.751 提升至 0.921）上均取得了显著性能提升，且在标准笔记本电脑 CPU 上单线程运行时间低于 150 毫秒。作者还透明地分享了成本明细、标注方法权衡（如规则衍生标签与音频大模型标注的对比）、编码器容量实验以及生产管道的局限性，所有数据、权重、代码及负面结果均已公开。

TamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech

TamilEOT: A Dataset and Model for Semantic End-of-Turn Detection in Tamil Telephone Speech

Summary

语音智能体常常难以准确判断用户何时结束发言，传统上依赖于固定的静音超时机制，这容易导致过早打断或响应迟钝。虽然针对各种语言的语义轮次结束检测器已经存在，但 TamilEOT 首次为南印度语种引入了此类开源资源。

Voice agents often struggle to determine precisely when a user has finished speaking, traditionally relying on fixed silence timeouts that can lead to premature interruptions or sluggish responses. While semantic end-of-turn detectors exist for various languages, TamilEOT introduces the first such open resource for a South Indian language.

该项目提供了： * TamilEOT 数据集： 从 116 场真实泰米尔语电话对话中提取的 18,485 个带标签的轮次边界。 * 纯音频检测器： 基于 Smart Turn v3 微调的两个模型，与零样本基线相比性能显着提升（准确率从 70.30% 提升至 83.71% 和 86.13%，ROC-AUC 从 0.751 提升至 0.921）。 * 高效性： 两个模型在标准笔记本电脑 CPU 上单线程运行时间均低于 150 毫秒。 * 透明的发现： 作者分享了全面的成本明细、标注权衡（对比规则衍生标签与音频大模型标注器）、编码器容量实验以及生产流水线的局限性。

The project delivers: * TamilEOT Dataset: 18,485 labeled turn boundaries extracted from 116 real Tamil telephone conversations. * Audio-Only Detectors: Two models fine-tuned from Smart Turn v3, demonstrating significant performance gains over zero-shot baselines (increasing accuracy from 70.30% to 83.71% and 86.13%, and ROC-AUC from 0.751 to 0.921). * Efficiency: Both models run in under 150 ms single-threaded on a standard laptop CPU. * Transparent Findings: The author shares comprehensive cost breakdowns, labeling trade-offs (comparing rule-derived labels against audio-LLM labellers), encoder capacity experiments, and production pipeline limitations.

Abstract

语音智能体在每一次停顿处都必须决定用户是否已经结束发言。如果没有语言模型，该决策就会退化为固定的静音超时：设置太短，智能体就会打断用户；设置太长，每一次轮次切换都要付出漫长的等待代价。开源的语义轮次结束检测器虽然存在，但据我们所知，没有一个涵盖南印度语种。我们发布了 TamilEOT：从 116 场真实的泰米尔语电话对话中截取了 18,485 个带有标签的轮次边界，以及两个基于 Smart Turn v3 微调的纯音频检测器。在包含来自 30 个未见通话的 4,168 个音频片段的独立测试集上，准确率从零样本的 70.30% 提升至 83.71%（8.7 MB）和 86.13%（21 MB）；ROC-AUC 从 0.751 提升至 0.921。两个模型在笔记本电脑 CPU 上单线程运行均可在 150 毫秒内完成。我们还报告了构建该系统的成本。通过盲测人工听辨校验，规则衍生的标签在正类上的准确率为 95.9%，但在负类上的准确率仅为 44.4%，低于随机猜测水平，这是因为规则所回答的问题与模型被要求回答的问题并不相同。将其替换为经测量与人工一致率达 97.5% 的音频大模型标注器，成本为 5.69 美元。在我们测量所有的训练杠杆中，只有编码器容量改变了结果；在相同配置和随机种子下的三次运行结果在 0.87 个准确率百分点内波动，这是我们其他所有变量微调都无法突破的下限。将相同的带标签边界通过生产环境的 VAD（语音活动检测器）和流式适配器回放，会额外损失 2.60 个百分点，并且有 7.8% 的边界根本不会呈现给模型。数据、权重、代码以及所有负面结果均已公开。

A voice agent has to decide, at every pause, whether the user has finished speaking. Without a model of the language that decision falls back to a fixed silence timeout: set it short and the agent interrupts, set it long and every turn pays the full wait. Open semantic end-of-turn detectors exist, but to our knowledge none covers a South Indian language. We release TamilEOT: 18,485 labelled turn boundaries cut from 116 real Tamil telephone conversations, and two audio-only detectors fine-tuned from Smart Turn v3. On a held-out split of 4,168 clips from 30 unseen calls, accuracy rises from 70.30% zero-shot to 83.71% (8.7 MB) and 86.13% (21 MB); ROC-AUC rises from 0.751 to 0.921. Both models run in under 150 ms single-threaded on a laptop CPU. We also report what building it cost. Rule-derived labels, checked against a blind human listening pass, were right 95.9% of the time on the positive class and 44.4% on the negative class, which is below chance, because the rule answered a different question than the model is asked. Replacing them with an audio-LLM labeller measured at 97.5% human agreement cost US$5.69. Of every training lever we measured, only encoder capacity moved the result; three runs at identical config and seed span 0.87 accuracy points, which is the floor below which none of our other deltas mean anything. Replaying the same labelled boundaries through the production VAD and streaming adapter costs a further 2.60 points, and 7.8% of boundaries are never surfaced to the model at all. Data, weights, code and every negative result are public.

Additional Metadata

Additional Metadata

license icon view license

license icon view license