SpokenNativQA: Multilingual Everyday Spoken Queries for LLMs
Firoj Alam, Md Arid Hasan, Shammur Absar Chowdhury
TL;DR
The paper presents SpokenNativQA, a multilingual spoken QA dataset designed to benchmark LLMs in realistic, everyday conversational settings. It provides ~33k samples in Arabic and English, with native speakers and culturally aligned topics, and includes a test subset drawn from real user data. The study benchmarks multiple ASR systems and LLMs under No-ASR and ASR conditions, using language-specific embeddings and a BERTScore-based evaluation. Findings show ASR errors significantly impact performance, though audio-enabled LLMs like GPT-4o-audio offer strong results, underscoring the need for cascade-less SQA systems for robust, real-world use.
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.