IndicContextEval

license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
language:
  - bn
  - gu
  - hi
  - ml
  - mr
  - or
  - te
  - ur
tags:
  - contextual-asr
  - audio-llm
  - speech-recognition
  - indic
  - multilingual
  - benchmark
pretty_name: IndicContextEval
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: Bengali
        path: data/Bengali-*
      - split: Gujarati
        path: data/Gujarati-*
      - split: Hindi
        path: data/Hindi-*
      - split: Malayalam
        path: data/Malayalam-*
      - split: Marathi
        path: data/Marathi-*
      - split: Odia
        path: data/Odia-*
      - split: Telugu
        path: data/Telugu-*
      - split: Urdu
        path: data/Urdu-*
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: reference_text_with_tags
      dtype: string
    - name: reference_text_no_tags
      dtype: string
    - name: reference_text_no_brackets
      dtype: string
    - name: language
      dtype: string
    - name: reference_text
      dtype: string
    - name: duration
      dtype: float64
    - name: task_name
      dtype: string
    - name: domain_name
      dtype: string
    - name: domain_description
      dtype: string
    - name: location
      dtype: string
    - name: speech_style
      dtype: string
    - name: code_mixing
      dtype: bool
    - name: audio_chunk_description
      dtype: string
    - name: l0_prompt
      dtype: string
    - name: l1_prompt
      dtype: string
    - name: l2_prompt
      dtype: string
    - name: l3_prompt
      dtype: string
    - name: l4_prompt
      dtype: string
    - name: l5_prompt
      dtype: string
    - name: l6_prompt
      dtype: string
    - name: audio_path
      dtype: string
    - name: speaker_id
      dtype: string
    - name: gender
      dtype: string
    - name: age
      dtype: string
    - name: file_name
      dtype: string
  splits:
    - name: Bengali
      num_bytes: 453299204
      num_examples: 1167
    - name: Gujarati
      num_bytes: 513084894
      num_examples: 1433
    - name: Hindi
      num_bytes: 1265929184
      num_examples: 3110
    - name: Malayalam
      num_bytes: 463355467
      num_examples: 1207
    - name: Marathi
      num_bytes: 1523363127
      num_examples: 3114
    - name: Odia
      num_bytes: 670758527
      num_examples: 1836
    - name: Telugu
      num_bytes: 1505064735
      num_examples: 4088
    - name: Urdu
      num_bytes: 393617253
      num_examples: 929
  download_size: 6481489663
  dataset_size: 6788472391

NOTE: This is a duplicate repo of "https://huggingface.co/datasets/ai4bharat/IndicContextEval" - visit the reference dataset - for any new updates made after Jul 30, 2026.

IndicContextEval

A Benchmark for Evaluating Context Utilisation in Audio Large Language Models Across 8 Indic Languages

Code and resources: https://github.com/AI4Bharat/IndicContextEval

Dataset at a glance


Languages	Hindi, Bengali, Telugu, Marathi, Gujarati, Malayalam, Odia, Urdu
Speakers	555
Duration	55.93 h
Utterances	16,884
Domains	23 professional domains
Speech styles	Read, Extempore
Prompt levels	L0–L6 (7 levels)

Per-language statistics

Language	Utterances	Hours	Speakers	Read	Extempore
Bengali	1,167	3.81	30	770	397
Gujarati	1,433	4.21	61	433	1,000
Hindi	3,110	9.08	124	1,237	1,873
Malayalam	1,207	4.03	15	826	381
Marathi	3,114	11.67	74	2,075	1,039
Odia	1,836	6.07	19	1,412	424
Telugu	4,088	13.70	205	1,425	2,663
Urdu	929	3.37	27	833	96
Total	16,884	55.93	555

Domains (23)

Academic Research & Publishing · Animal Sciences · Architecture & Urban Planning · Arts · Audio & Media Technology · Business · Core Engineering · Culinary Arts & Food Science · Data Science · Defense & Armed Forces · Film & Media Production · Forensics & Legal Sciences · Fundamental Sciences · Gaming & Media · Humanities · Linguistics · Materials Science · Medical Sciences · Robotics & Automation Engineering · Skilled Trades · Social Sciences · Supply Chain & Logistics · Textiles & Fashion

Prompt taxonomy (L0–L6)

Each level adds exactly one contextual signal; everything else is held constant. Output is always required in the native script of the target language.

Level	Context added
L0	None — bare transcription instruction, no language hint
L1	Target language specified (baseline)
L2	+ structured domain metadata (style, region, one-line domain description)
L3	+ natural-language audio description
L4	+ domain entity list in English script
L5	+ the same domain entity list in native script
L6	+ an entity list from an unrelated domain (adversarial control)

Loading

Each language is a split (with playable, embedded audio):

from datasets import load_dataset

ds = load_dataset("SakshiJ/IndicContextEval", split="Hindi")   # Bengali, Telugu, ...
print(ds[0]["audio"], ds[0]["l5_prompt"])                       # audio decoded automatically

License

Released under CC-BY-4.0. You may share and adapt the material for any purpose, including commercially, provided you give appropriate credit.

Citation

@misc{joshi2026indiccontextevalbenchmarkevaluatingcontext,
      title={IndicContextEval: A Benchmark for Evaluating Context Utilisation in Audio Large Language Models Across 8 Indic Languages},
      author={Sakshi Joshi and Dhruv Subhash Rathi and Sanskar Singh and Eldho Ittan George and R J Hari and Kaushal Bhogale and Mitesh M. Khapra},
      year={2026},
      eprint={2606.19157},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2606.19157},
}