Spelling-Aware Word-Based End-to-End ASR

We propose a new end-to-end architecture for automaticspeech recognition that expands the listen, attend andspell (LAS) paradigm. While the main word-predicting networkis trained to predict words, the secondary, speller network, isoptimized to predict word spellings from inner representationsof the main network (e.g. word embeddings or context vectorsfrom the attention module). We show that this joint trainingimproves the word error rate of a word-based system and enablessolving additional tasks, such as out-of-vocabulary word detectionand recovery. The tests are conducted on LibriSpeech datasetconsisting of 1000h of read speech.