Relaxed Attention for Transformer Models

The powerful modeling capabilities of all-attention-based transformer architectures often cause overfitting and - for natural language processing tasks - lead to an implicitly learned internal language model in the autoregressive transformer decoder complicating the integration of external language models. In this paper, we explore relaxed attention, a simple and easy-to-implement smoothing of the attention weights, yielding a two-fold improvement to the general transformer architecture: First, relaxed attention provides regularization when applied to the self-attention layers in the encoder. Second, we show that it naturally supports the integration of an external language model as it suppresses the implicitly learned internal language model by relaxing the cross attention in the decoder. We demonstrate the benefit of relaxed attention across several tasks with clear improvement in combination with recent benchmark approaches. Specifically, we exceed the former state-of-the-art performance of 26.90% word error rate on the largest public lip-reading LRS3 benchmark with a word error rate of 26.31%, as well as we achieve a top-performing BLEU score of 37.67 on the IWSLT14 (DE$\rightarrow$EN) machine translation task without external language models and virtually no additional model parameters. Code and models will be made publicly available.

Relaxed Attention: A Simple Method to Boost Performance of End-to-End Automatic Speech Recognition

Recently, attention-based encoder-decoder (AED) models have shown high performance for end-to-end automatic speech recognition (ASR) across several tasks. Addressing overconfidence in such models, in this paper we introduce the concept of relaxed attention, which is a simple gradual injection of a uniform distribution to the encoder-decoder attention weights during training that is easily implemented with two lines of code. We investigate the effect of relaxed attention across different AED model architectures and two prominent ASR tasks, Wall Street Journal (WSJ) and Librispeech. We found that transformers trained with relaxed attention outperform the standard baseline models consistently during decoding with external language models. On WSJ, we set a new benchmark for transformer-based end-to-end speech recognition with a word error rate of 3.65%, outperforming state of the art (4.20%) by 13.1% relative, while introducing only a single hyperparameter. Upon acceptance, models will be published on github.

Multi-Encoder Learning and Stream Fusion for Transformer-Based End-to-End Automatic Speech Recognition

Stream fusion, also known as system combination, is a common technique in automatic speech recognition for traditional hybrid hidden Markov model approaches, yet mostly unexplored for modern deep neural network end-to-end model architectures. Here, we investigate various fusion techniques for the all-attention-based encoder-decoder architecture known as the transformer, striving to achieve optimal fusion by investigating different fusion levels in an example single-microphone setting with fusion of standard magnitude and phase features. We introduce a novel multi-encoder learning method that performs a weighted combination of two encoder-decoder multi-head attention outputs only during training. Employing then only the magnitude feature encoder in inference, we are able to show consistent improvement on Wall Street Journal (WSJ) with language model and on Librispeech, without increase in runtime or parameters. Combining two such multi-encoder trained models by a simple late fusion in inference, we achieve state-of-the-art performance for transformer-based models on WSJ with a significant WER reduction of 19\% relative compared to the current benchmark approach.