STOP: A dataset for Spoken Task Oriented Semantic Parsing

TL;DR

This paper introduces the STOP dataset, a large and complex spoken language understanding dataset with semantic parse labels, to facilitate research in end-to-end SLU models that directly predict intent from audio.

Contribution

It releases the largest publicly available SLU dataset with human and TTS audio, and establishes low-resource benchmarks for end-to-end SLU system development.

Findings

End-to-end SLU models perform slightly worse than cascaded models.

The dataset includes human-recorded and TTS-generated audio for domain adaptation.

Benchmark results highlight challenges and opportunities for low-resource SLU.

Abstract

End-to-end spoken language understanding (SLU) predicts intent directly from audio using a single model. It promises to improve the performance of assistant systems by leveraging acoustic information lost in the intermediate textual representation and preventing cascading errors from Automatic Speech Recognition (ASR). Further, having one unified model has efficiency advantages when deploying assistant systems on-device. However, the limited number of public audio datasets with semantic parse labels hinders the research progress in this area. In this paper, we release the Spoken Task-Oriented semantic Parsing (STOP) dataset, the largest and most complex SLU dataset to be publicly available. Additionally, we define low-resource splits to establish a benchmark for improving SLU when limited labeled data is available. Furthermore, in addition to the human-recorded audio, we are releasing a TTS-generated version to benchmark the performance for low-resource domain adaptation of end-to-end SLU systems. Initial experimentation show end-to-end SLU models performing slightly worse than their cascaded counterparts, which we hope encourages future work in this direction.