Exploring end-to-end framework towards Khasi speech recognition system

摘要

Building a conventional automatic speech recognition (ASR) system based on hidden Markov model (HMM)/deep neural network (DNN) makes the system complex as it requires various modules such as acoustic, lexicon, linguistic resources, language models etc. particularly with the low resource languages. In contrast, End-to-End architecture has greatly simplifies the model building process by representing complex modules with a simple deep network and by replacing the use of linguistic resources with a data-driven learning techniques. In this paper, we present our prior work by exploring End-to-End (E2E) framework for Khasi speech recognition system and the novel extension towards the development of speech corpora for standard Khasi dialect. We implemented the proposed E2E model by using Nabu ASR toolkit. Additionally, three other models (monophone, triphone and hybrid DNN) were built. Comparing the results, significant improvement was achieved using the proposed method particularly with the connectionist temporal classification (CTC) with a character error rate (CER) of 5.04%.