🗣️ SqCLIRIL: Spoken Query Benchmark for Cross-Lingual IR in Indian Languages

SqCLIRIL is a Spoken Query Benchmark designed to evaluate cross-lingual information retrieval (CLIR) systems using both spoken and text queries.
It covers five Indian languages — Hindi, Gujarati, Bengali, Kannada, and English — with diverse speech samples from male and female speakers to capture natural variability in pronunciation and acoustic conditions.

---

📘 Dataset Summary

Feature	Description
Dataset name	SqCLIRIL
Languages	Hindi (hi), Gujarati (gu), Bengali (bn), Kannada (kn), English (en)
Modalities	Text, Speech (WAV), ASR Transcriptions
Speakers	Male and Female
Data Type	Queries and their spoken utterances
Format	.tsv (queries and transcriptions), .wav (spoken queries)

---

🧩 Dataset Structure

The dataset is organized into three main folders:

SqCLIRIL/
│
├── text/
│   ├── hindi/
│   │   ├── trec_dl19_hindi_query.tsv
│   │   ├── trec_dl20_hindi_query.tsv
│   │   └── trec_dl1920_hindi_query.tsv
│   ├── gujarati/
│   ├── bengali/
│   ├── kannada/
│   └── english/
│
├── asr/
│   ├── hindi/
│   │   ├── male/
│   │   │   ├── m1/ → sq_hi_m1.tsv
│   │   │   ├── m2/ → sq_hi_m2.tsv
│   │   └── female/
│   │       ├── f1/ → sq_hi_f1.tsv
│   │       ├── f2/ → sq_hi_f2.tsv
│   ├── ...
│
└── speech/
    ├── hindi/
    │   ├── male/
    │   │   ├── m1/ → 123.wav, 124.wav, ...
    │   │   ├── m2/ → ...
    │   └── female/
    │       ├── f1/ → 201.wav, ...
    │       ├── f2/ → ...
    ├── ...

Folder Descriptions

• text/: Contains text queries for each language in three benchmark splits (trec_dl1920, trec_dl19, trec_dl20).
• speech/: Contains recorded spoken queries (WAV files) from both male and female speakers.
• asr/: Contains automatic speech recognition (ASR) transcriptions of the spoken queries, structured by gender and speaker ID.

---

🗂️ Example Structure (Hindi)

SqCLIRIL/
├── text/hindi/trec_dl19_hindi_query.tsv
├── speech/hindi/male/m1/123.wav
├── speech/hindi/male/m1/124.wav
├── speech/hindi/female/f1/201.wav
├── asr/hindi/male/m1/sq_hi_m1.tsv
├── asr/hindi/female/f1/sq_hi_f1.tsv

Each line in the sq_hi_f1.tsv corresponds to the transcription of the spoken file with the same query ID (e.g., 123.wav).

---

💡 Intended Uses

• Cross-lingual information retrieval (CLIR)
• Speech-to-text retrieval
• Multilingual query understanding
• Spoken Query Search in Indian Languages

---

⚙️ Data Fields

Field	Description
query_id	Unique identifier for the query (e.g., 123)
language	One of {hi, gu, bn, kn, en}
text_query	Original text form of the query
speech_audio	Path to the .wav file containing the spoken version
asr_transcription	Automatic transcription of the spoken query
speaker_id	Speaker identity (e.g., m1, f2)
gender	Male/Female

---

📊 Data Splits

Each language contains three splits:

Split	Description
trec_dl19	TREC Deep Learning Track 2019 queries
trec_dl20	TREC Deep Learning Track 2020 queries
trec_dl1920	Combined 2019–2020 queries

---

🎧 Audio Details

Property	Value
Format	WAV
Sampling Rate	16 kHz
Channels	Mono
Environment	Natural home and lab settings

---

📜 Citation

If you use this dataset, please cite:

@article{DAVE2025,
  title = {SqCLIRIL: Spoken query cross-lingual information retrieval in Indian languages},
  journal = {Pattern Recognition Letters},
  year = {2025},
  issn = {0167-8655},
  doi = {https://doi.org/10.1016/j.patrec.2025.08.022},
  url = {https://www.sciencedirect.com/science/article/pii/S0167865525003071},
  author = {Bhargav Dave and Prasenjit Majumder},
}

---

⚖️ License

License: CC BY 4.0

---

🙌 Acknowledgements

We thank all contributors and speakers involved in building this multilingual benchmark for advancing speech-based cross-lingual IR research in India.