Showing 1–1 of 1 results for author: Dinh, T Q K

A language score based output selection method for multilingual speech recognition

Authors: Van Huy Nguyen, Thi Quynh Khanh Dinh, Truong Thinh Nguyen, Dang Khoa Mac

Abstract: The quality of a multilingual speech recognition system can be improved by adaptation methods if the input language is specified. For systems that can accept multilingual inputs, the popular approach is to apply a language identifier to the input then switch or configure decoders in the next step, or use one more subsequence model to select the output from a set of candidates. Motivated by the goa… ▽ More The quality of a multilingual speech recognition system can be improved by adaptation methods if the input language is specified. For systems that can accept multilingual inputs, the popular approach is to apply a language identifier to the input then switch or configure decoders in the next step, or use one more subsequence model to select the output from a set of candidates. Motivated by the goal of reducing the latency for real-time applications, in this paper, a language model rescoring method is firstly applied to produce all possible candidates for target languages, then a simple score is proposed to automatically select the output without any identifier model or language specification of the input language. The main point is that this score can be simply and automatically estimated on-the-fly so that the whole decoding pipeline is more simple and compact. Experimental results showed that this method can achieve the same quality as when the input language is specified. In addition, we present to design an English and Vietnamese End-to-End model to deal with not only the problem of cross-lingual speakers but also as a solution to improve the accuracy of borrowed words of English in Vietnamese. △ Less

Submitted 2 May, 2020; originally announced May 2020.