| --- |
| license: cc-by-sa-4.0 |
| task_categories: |
| - text-retrieval |
| - feature-extraction |
| - automatic-speech-recognition |
| language: |
| - en |
| - zh |
| tags: |
| - spoken-query-retrieval |
| - information-retrieval |
| - audio-text-retrieval |
| - mteb |
| - audio |
| - c-mteb |
| - robustness |
| pretty_name: SQuTR |
| size_categories: |
| - 10K<n<100K |
| --- |
| # 🏆 Recognition |
| * **[2026-05]** SQuTR📄 was ACCEPTED to **SIGIR 2026**! 🎉 |
| * **[2026-02]** SQuTR was featured as the **#1 Paper of the Day** on [Hugging Face Daily Papers](https://huggingface.co/papers/2602.12783)! |
|
|
|
|
| # SQuTR: A Robustness Benchmark for Spoken Query to Text Retrieval |
|
|
| [](https://huggingface.co/papers/2602.12783) |
|
|
| [](https://github.com/ttoyekk1a/SQuTR-Spoken-Query-to-Text-Retrieval) |
| [](https://arxiv.org/abs/2602.12783) |
|
|
| **SQuTR** (Spoken Query-to-Text Retrieval) is a large-scale bilingual benchmark designed to evaluate the robustness of information retrieval systems under realistic acoustic perturbations. |
|
|
| While speech interaction is becoming a primary interface for IR systems, performance often degrades significantly in noisy environments. SQuTR provides a standardized framework featuring **37,317** complex queries across **6 domains**, synthesized with **200 real speakers**, and evaluated under **4 graded noise levels**. |
|
|
| --- |
|
|
| ## 🌟 Key Features |
|
|
| * **Bilingual & Multi-Domain:** Includes 6 subsets from MTEB and C-MTEB covering Wikipedia, Finance, Medical, and Encyclopedia domains. |
| * **High-Fidelity Synthesis:** Generated using **CosyVoice-3** with diverse speaker profiles, totaling **190.4 hours** of audio. |
| * **Robustness Evaluation:** Explicitly models four acoustic conditions: **Clean, Low Noise (20dB), Medium Noise (10dB), and High Noise (0dB)**. |
| * **MTEB Compatibility:** Follows standard JSONL/BEIR formatting for seamless integration into modern retrieval pipelines. |
|
|
| --- |
|
|
| ## 📂 Dataset Structure |
|
|
| The dataset is organized by language and subset. Each subset (e.g., `fiqa`) contains the original text documents and the synthesized audio queries under different SNR conditions. |
|
|
| ```text |
| SQuTR/ |
| └── source_data/ |
| ├── en/ (English Datasets: fiqa, hotpotqa, nq) |
| │ └── [subset_name]/ |
| │ ├── audio_clean/ # Clean original audio files (.wav) |
| │ ├── audio_noise_snr_0/ # Audio with 0dB Signal-to-Noise Ratio |
| │ ├── audio_noise_snr_10/ # Audio with 10dB Signal-to-Noise Ratio |
| │ ├── audio_noise_snr_20/ # Audio with 20dB Signal-to-Noise Ratio |
| │ ├── qrels/ # Query relevance judgments (TSV/JSONL) |
| │ ├── corpus.jsonl # Text corpus documents |
| │ ├── queries.jsonl # Original text queries |
| │ ├── queries_with_audio_clean.jsonl # Metadata mapping text to clean audio |
| │ ├── queries_with_audio_noise_snr_0.jsonl # Metadata for 0dB noise queries |
| │ ├── queries_with_audio_noise_snr_10.jsonl # Metadata for 10dB noise queries |
| │ └── queries_with_audio_noise_snr_20.jsonl # Metadata for 20dB noise queries |
| └── zh/ (Chinese Datasets: DuRetrieval, MedicalRetrieval, T2Retrieval) |
| └── [subset_name]/ |
| └── (Same structure as above) |
| ``` |
| --- |
|
|
| ## 💾 How to Use the Dataset |
|
|
| You can download the dataset directly from this Hugging Face repository. To use the evaluation scripts, please refer to our [GitHub Repository](https://github.com/ttoyekk1a/SQuTR-Spoken-Query-to-Text-Retrieval). |
