Leveraging Unpaired Text Data for Training End-to-End Speech-to-Intent Systems

TL;DR

This paper explores leveraging unpaired text data and transfer learning techniques to improve end-to-end speech-to-intent systems, reducing the need for large amounts of labeled speech data while maintaining high accuracy.

Contribution

It introduces methods to incorporate text resources into speech-to-intent models, including transfer learning with BERT embeddings and data augmentation via text-to-speech, achieving significant performance recovery.

Findings

Matching state-of-the-art performance with less data

Transfer learning improves intent classification accuracy

Data augmentation recovers 80% of performance loss

Abstract

Training an end-to-end (E2E) neural network speech-to-intent (S2I) system that directly extracts intents from speech requires large amounts of intent-labeled speech data, which is time consuming and expensive to collect. Initializing the S2I model with an ASR model trained on copious speech data can alleviate data sparsity. In this paper, we attempt to leverage NLU text resources. We implemented a CTC-based S2I system that matches the performance of a state-of-the-art, traditional cascaded SLU system. We performed controlled experiments with varying amounts of speech and text training data. When only a tenth of the original data is available, intent classification accuracy degrades by 7.6% absolute. Assuming we have additional text-to-intent data (without speech) available, we investigated two techniques to improve the S2I system: (1) transfer learning, in which acoustic embeddings for intent classification are tied to fine-tuned BERT text embeddings; and (2) data augmentation, in which the text-to-intent data is converted into speech-to-intent data using a multi-speaker text-to-speech system. The proposed approaches recover 80% of performance lost due to using limited intent-labeled speech.