Attentionless Streaming ASR

attentionless-streaming-asr contains the codebase associated with the paper Do we really need Self-Attention for Streaming Automatic Speech Recognition?, which investigates the role of self-attention mechanisms in streaming automatic speech recognition (ASR).

The code release aims to promote transparency, reproducibility, and further innovation within the speech recognition community. It includes a minimalist implementation of models based on the Conformer-Transducer architecture, along with training scripts, configuration files, and detailed documentation. The datasets used for the experiments, such as LibriSpeech and TEDLIUM, are publicly available, ensuring that other researchers can replicate the results and build upon this work, with the goal of encouraging collaborative development of more efficient streaming ASR models, especially those, that reduce computational costs, by replacing, or removing, self-attention modules.

The code is available on GitHub under the MIT licence.