SpokenNativQA: Multilingual Everyday Spoken Queries for LLMs

TL;DR

SpokenNativQA is a comprehensive multilingual spoken question-answering dataset that evaluates LLMs in real-world conversational speech scenarios, including low-resource languages and dialects, addressing a gap in speech-based AI benchmarking.

Contribution

We introduce SpokenNativQA, the first multilingual, culturally aligned spoken QA dataset with speech variability, and benchmark LLMs and ASR systems on it.

Findings

Benchmark results highlight challenges in low-resource language understanding.

ASR accuracy varies significantly across languages and dialects.

LLMs show promising performance but need improvements for speech-based multilingual QA.

Abstract

Large Language Models (LLMs) have demonstrated remarkable performance across various disciplines and tasks. However, benchmarking their capabilities with multilingual spoken queries remains largely unexplored. In this study, we introduce SpokenNativQA, the first multilingual and culturally aligned spoken question-answering (SQA) dataset designed to evaluate LLMs in real-world conversational settings. The dataset comprises approximately 33,000 naturally spoken questions and answers in multiple languages, including low-resource and dialect-rich languages, providing a robust benchmark for assessing LLM performance in speech-based interactions. SpokenNativQA addresses the limitations of text-based QA datasets by incorporating speech variability, accents, and linguistic diversity. We benchmark different ASR systems and LLMs for SQA and present our findings. We released the data at (https://huggingface.co/datasets/QCRI/SpokenNativQA) and the experimental scripts at (https://llmebench.qcri.org/) for the research community.