Spoken QA dataset

SpokenNativQA: Multilingual Everyday Spoken Queries for LLMs

SpokenNativQA is a multilingual and culturally aligned spoken question-answering dataset for evaluating LLMs in real-world conversational settings, where speech variability, accents, ASR errors, and local information needs all affect model behavior.

Hugging Face dataset Paper PDF Back to resources

Scope Everyday spoken QA with ASR transcriptions, answers, and culturally grounded queries.

Overview

SpokenNativQA extends culturally grounded question answering into spoken interaction. The resource is designed for speech-centric evaluation rather than text-only QA, so it captures how ASR systems and LLMs behave when everyday questions are asked naturally and then evaluated with the corresponding answers.

The Interspeech 2025 paper describes the dataset as the first multilingual and culturally aligned spoken QA resource for LLM evaluation in real-world conversational settings, with approximately 33,000 naturally spoken questions and answers. The public Hugging Face dataset currently exposes Arabic and English resources with multiple ASR transcription variants.

Dataset at a glance

150

Hugging Face downloads

Hugging Face likes

33k

spoken questions and answers reported in the paper

public ASR dataset subsets

languages tagged on Hugging Face

ASR systems represented

Audio + text

modalities

CC BY-NC-SA 4.0

license

What the resource includes

Natural spoken questions

The dataset focuses on everyday spoken queries, preserving speech variation and local context that are often missing from text-only QA benchmarks.

ASR transcription variants

The public configs include Arabic and English subsets for Azure, Fanar Aura, Google, and Whisper ASR outputs, enabling direct study of recognition noise.

Answer-grounded evaluation

Rows include language, audio file metadata, question text, answer text, location, and ASR transcription fields for evaluating speech-to-answer pipelines.

Dataset fields

The Hugging Face viewer shows fields for language, dataset id, audio file name and path, the original question, answer, location, and question ASR transcription. This makes the resource useful for benchmarking both ASR systems and downstream LLM answer quality.

Question answering Audio Text Arabic English ASR robustness Cultural alignment

Citation

@inproceedings{alam25_interspeech,
  title     = {{SpokenNativQA: Multilingual Everyday Spoken Queries for LLMs}},
  author    = {Firoj Alam and Md Arid Hasan and Shammur Absar Chowdhury},
  year      = {2025},
  booktitle = {{Interspeech 2025}},
  pages     = {2685--2689},
  doi       = {10.21437/Interspeech.2025-2011},
  issn      = {2958-1796},
}

Metrics are fetched from public APIs when available. Last checked: runtime.