Transformer Model Compression for End-to-End Speech Recognition on Mobile Devices

Transformer-based models have achieved state-of-the-art performance in various areas of machine learning, including automatic speech recognition. However, their cost in terms of computational power, memory or energy consumption can be exorbitant, hence the interest in compression techniques. Trans-former models are mostly composed of attention and feedforward components. In this paper, we propose to reduce the size of a transformer model in an end-to-end speech recognition system by decreasing the number and precision of linear layer parameters. Specifically, we investigate the impact of weight pruning on system performance. We then consider model quantization. To further reduce the model size, we address the combination of pruning and quantization methods. Experiments carried out on several speech datasets from different languages show that the memory footprint can be reduced by up to 84% with an insignificant loss of accuracy.