# ViSQA: A benchmark dataset and baseline models for Vietnamese spoken question answering

## TL;DR

ViSQA is the first benchmark for Vietnamese spoken question answering, enabling evaluation of models under varying transcription quality.

## Contribution

ViSQA introduces the first standardized Vietnamese SQA benchmark with over 13,000 question-answer pairs and audio variants.

## Key findings

- ASR errors significantly degrade model performance, e.g., ViT5 EM drops from 62.04% to 36.30%.
- Training on spoken transcriptions improves robustness, raising ViT5 EM from 36.30% to 50.70%.
- ViSQA enables systematic analysis of ASR error impact on downstream reasoning.

## Abstract

Spoken Question Answering (SQA) extends machine reading comprehension to spoken content and requires models to handle both automatic speech recognition (ASR) errors and downstream language understanding. Although large-scale SQA benchmarks exist for high-resource languages, Vietnamese remains underexplored due to the lack of standardized datasets. This paper introduces ViSQA, the first benchmark for Vietnamese Spoken Question Answering. ViSQA extends the UIT-ViQuAD corpus using a reproducible text-to-speech and ASR pipeline, resulting in over 13,000 question–answer pairs aligned with spoken inputs. The dataset includes clean and noise-degraded audio variants to enable systematic evaluation under varying transcription quality. Experiments with five transformer-based models show that ASR errors substantially degrade performance (e.g., ViT5 EM: 62.04% → 36.30%), while training on spoken transcriptions improves robustness (ViT5 EM: 36.30% → 50.70%). ViSQA provides a rigorous benchmark for evaluating Vietnamese SQA systems and enables systematic analysis of the impact of ASR errors on downstream reasoning.