RNN-Transducer-based Losses for Speech Recognition on Noisy Targets

We start with the template "CM3015 Machine Learning and Neural Networks, Theme 1: Deep Learning on a public dataset," which describes the task of choosing a publicly available dataset and training a deep learning model on it. So, we will work with a neural network-based end-to-end ASR system, using LibriSpeech [2] dataset, a popular academic benchmark. We limit our task to RNN-T ransducer [3] systems, which are widely used in production and provide state-of-the-art quality [4] in most cases. We are going beyond the standard task and focusing our research on making RNN-T ransducer systems robust to noisy targets: unlike well-curated datasets, in the industry, the training data contains different errors due to the unreliability of the transcription sources or the inability to transcribe noisy speech accurately . T o solve the problem of training on the noisy data, we will analyze the impact of different types of errors in training data on the quality of the RNN-T ransducer system and explore different loss modifications to overcome the problem. We will construct the artificial training data by mutating correct transcripts from the LibriSpeech [2] training part, similar to the approaches used in the related work, and try to achieve the best possible quality on the development and test data standard for LibriSpeech.