Alternate Intermediate Conditioning with Syllable-level and Character-level Targets for Japanese ASR

However, the mapping can be Although end-to-end ASR methods achieved sufficient problematic when several different pronunciations should performance in English, the challenge remains in languages be mapped into one character or when one pronunciation such as Japanese, which face a large character vocabulary is shared among many different characters. Japanese ASR size with many homophones and multiple pronunciations suffers the most from such many-to-one and one-to-many [14]. The character vocabulary size is larger than that of mapping problems due to Japanese kanji characters. To alleviate phonogram languages such as English. Japanese has over the problems, we introduce explicit interaction between three thousand characters, while English has at most about characters and syllables using Self-conditioned connectionist one hundred characters. Japanese ASR also suffers from homophones: temporal classification (CTC), in which the upper layers are many characters share the same pronunciation, "self-conditioned" on the intermediate predictions from the e.g., 高, 公, 行 and other hundreds of characters have the lower layers. The proposed method utilizes character-level same pronunciation "kou". Therefore, an acoustic feature and syllable-level intermediate predictions as conditioning should be mapped to different character labels considering features to deal with mutual dependency between characters language contexts.