Pretrained Semantic Speech Embeddings for End-to-End Spoken Language Understanding via Cross-Modal Teacher-Student Learning

TL;DR

This paper introduces a cross-modal teacher-student training approach that leverages pretrained speech and text embeddings to develop end-to-end spoken language understanding systems, reducing error propagation and improving performance.

Contribution

It presents a novel method combining pretrained speech recognition and semantic embeddings within a teacher-student framework for end-to-end spoken language understanding.

Findings

Achieves comparable performance to pipeline systems without training data.

Outperforms pipeline systems after minimal fine-tuning on two benchmarks.

Demonstrates effectiveness across three benchmark datasets.

Abstract

Spoken language understanding is typically based on pipeline architectures including speech recognition and natural language understanding steps. These components are optimized independently to allow usage of available data, but the overall system suffers from error propagation. In this paper, we propose a novel training method that enables pretrained contextual embeddings to process acoustic features. In particular, we extend it with an encoder of pretrained speech recognition systems in order to construct end-to-end spoken language understanding systems. Our proposed method is based on the teacher-student framework across speech and text modalities that aligns the acoustic and the semantic latent spaces. Experimental results in three benchmarks show that our system reaches the performance comparable to the pipeline architecture without using any training data and outperforms it after fine-tuning with ten examples per class on two out of three benchmarks.