IndicContextEval
收藏资源简介:
IndicContextEval是由AI4Bharat和Sarvam AI联合创建的多语言语音基准数据集,旨在系统评估音频大语言模型在上下文利用方面的性能。该数据集包含56小时的自然语音,覆盖8种印度语言(印地语、孟加拉语等)和23个专业领域,数据来源于555名说话者的朗读和即兴演讲,并经过严格的质量控制和人工转录。数据集通过设计七级提示框架,逐步引入元数据、自然语言描述和实体列表等上下文信号,以分析模型对特定语境信息的依赖程度。其核心应用在于评测和提升音频大语言模型的上下文感知能力,解决传统语音识别系统在利用提示信息时可能存在的参数记忆而非真正语境理解的问题。
IndicContextEval is a multilingual speech benchmark dataset jointly created by AI4Bharat and Sarvam AI, which is designed to systematically evaluate the context utilization performance of audio large language models (LLMs). This dataset comprises 56 hours of natural speech, covering 8 Indian languages (Hindi, Bengali, etc.) and 23 professional domains. The data is sourced from read and impromptu speeches delivered by 555 speakers, and has undergone rigorous quality control and manual transcription. The dataset designs a seven-level prompt framework to gradually introduce contextual signals such as metadata, natural language descriptions and entity lists, so as to analyze the model's dependence on specific contextual information. Its core application lies in evaluating and enhancing the contextual awareness capability of audio LLMs, and addressing the issue that traditional speech recognition systems may rely on parameter memory rather than genuine contextual understanding when utilizing prompt information.
IndicContextEval 数据集概述
基本信息
- 全称: IndicContextEval: A Benchmark for Evaluating Context Utilisation in Audio Large Language Models Across 8 Indic Languages
- 发布机构: AI4Bharat, IIT Madras 与 Sarvam AI
- 论文链接: https://arxiv.org/abs/2606.19157
- 数据集地址: https://huggingface.co/datasets/AI4Bharat/IndicContextEval
- 许可协议: CC BY 4.0
- 收录会议: Interspeech 2026
数据集规模
- 总时长: 56小时
- 语种覆盖: 8种印度语言(包括 Hindi、Bengali、Telugu 等)
- 说话人数: 555位
- 专业领域: 23个专业领域
数据集用途
评估音频大语言模型(AudioLLMs)在语音识别任务中对上下文信息的利用能力,特别是区分模型是否真正利用提供的上下文提示,还是依赖预训练阶段学到的参数化知识。
提示层级框架(L0–L6)
数据集设计了7级提示,逐步引入上下文信息:
| 层级 | 添加的上下文 |
|---|---|
| L0 | 无上下文(纯转录,无语言提示) |
| L1 | 指定目标语言(基线) |
| L2 | + 结构化领域元数据 |
| L3 | + 自然语言音频描述 |
| L4 | + 英文脚本实体列表 |
| L5 | + 本地语种脚本实体列表 |
| L6 | + 无关领域实体列表(对抗性) |
数据集构成
每个音频样本提供三种参考转录变体:
reference_text_with_tags: 本地语种 + 英文标签(原始)reference_text_no_tags: 仅本地语种(用于计算WER)reference_text_no_brackets: 本地语种 + 英文内联,无括号
评估指标
- WER(词错误率): 经Indic NLP Library归一化,并对每个样本进行200%参考长度的截断处理(防止解码失控/重复循环主导平均分)
- NEER(命名实体错误率): 参考命名实体中缺失或错误的比例
评估结果(截断WER%按提示层级)
| 层级 | GPT-4o | Gemini 3 Flash | Gemma 3n | Sarvam Audio |
|---|---|---|---|---|
| L0 | 29.83 | 24.30 | 51.21 | 20.39 |
| L1 | 28.61 | 18.90 | 38.73 | 16.86 |
| L2 | 28.37 | 19.28 | 52.20 | 16.78 |
| L3 | 26.08 | 18.39 | 40.22 | 16.43 |
| L4 | 27.97 | 19.88 | 46.37 | 16.80 |
| L5 | 26.04 | 17.46 | 43.11 | 15.70 |
| L6 | 28.47 | 19.67 | 47.95 | 16.69 |
独立ASR基线 IndicConformer (L1) 的 WER 为 18.81%,NEER 为 29.58%。
引用信息
bibtex @inproceedings{joshi2026indiccontexteval, title = {IndicContextEval: A Benchmark for Evaluating Context Utilisation in Audio Large Language Models Across 8 Indic Languages}, author = {Joshi, Sakshi and Rathi, Dhruv Subhash and Singh, Sanskar and George, Eldho Ittan and Hari, R J and Bhogale, Kaushal and Khapra, Mitesh M.}, booktitle = {Proc. Interspeech}, year = {2026}, note = {arXiv:2606.19157} }
联系方式
- 邮箱: sakshijcom@gmail.com





