Streaming End-to-End ASR Using CTC Decoder and DRA for Linguistic Information Substitution

This study proposes a method of streaming automatic speech recognition (ASR) for Japanese speech using domain adaptation. The development of ASR technology has led to its employment in various fields, however several issues, such as the decline recognition accuracy which occurs due to domain mismatches between training and test data. In particular, the accuracy of online speech recognition, also known as streaming ASR, which is becoming popular in real-world applications, also suffers from this issue. To realize domain-adapted streaming ASR, in this study we utilize a method of domain adaptation known as the Density Ratio Adaptation (DRA), which uses Language Models (LMs) trained with extensive text data. Our proposed method employs DRA for streaming ASR of Japanese speech, in a model using a CTC decoder for recognition. To facilitate streaming ASR processing, the CTC decoder successively replaces linguistic information on a frame-by-frame basis, obtaining recognition results through greedy searches. Considering CTC’s assumption of conditional independence, we used unigram LMs for the subtraction and addition of language information when using DRA, during which selected linguistic information is replaced. Experimental results demonstrate that the proposed method enhances speech recognition accuracy when there is data domain mismatch.