Improving End-to-End Bangla Speech Recognition with Semi-supervised Training

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1875–1883 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1875 Improving End-to-End Bangla Speech Recognition with Semi-supervised Training Nafis Sadeq∗2, Nafis Tahmid Chowdhury∗1, Farhan Tanvir Utshaw 1, Shafayat Ahmed1, and Muhammad Abdullah Adnan1 1Bangladesh University of Engineering and Technology (BUET) 2Samsung R&D Institute Bangladesh {1205008.ns,1505077.ntc,1505105.ftu,1205115.sa}@ugrad.cse.buet.ac.bd adnan@cse.buet.ac.bd Abstract Automatic speech recognition systems usually require large annotated speech corpus for train- ing. The manual annotation of a large corpus is very difficult. It can be very helpful to use un- supervised and semi-supervised learning meth- ods in addition to supervised learning. In this work, we focus on using a semi-supervised training approach for Bangla Speech Recogni- tion that can exploit large unpaired audio and text data. We encode speech and text data in an intermediate domain and propose a novel loss function based on the global encoding dis- tance between encoded data to guide the semi- supervised training. Our proposed method re- duces the Word Error Rate (WER) of the sys- tem from 37% to 31.9%. 1 Introduction 1 An annotated speech corpus is an essential com- ponent for the development of an automatic speech recognition system (ASR). Speech corpus is a col- lection of audio files with corresponding text tran- scriptions. Manually developing a speech corpus of required size is a time consuming and monotonous task. It also requires some prerequisites like a recording environment, clear utterance, and addi- tional information such as gender of speakers, etc. For achieving a large vocabulary continuous speech recognition we need approximately several hun- dred to few thousands of hours of speech corpus. Semi-supervised training can be a useful solution to tackle the hurdles related to speech corpus de- velopment. Semi-supervised training can provide us a way to exploit a huge collection of publicly available text as well as audio resources to improve the performance of an ASR. In this work, we focus on improving an end- to-end speech recognition system for Bangladeshi 1∗ authors contributed equally Bangla using semi-supervised training. There are very few publicly available large speech corpora for Bangladeshi Bangla. Google released 229 hours of speech corpus for Bangladeshi Bangla (Kjartansson et al., 2018). But there are huge amounts of pub- licly available news audio files, audiobooks, record- ings in Youtube and other media sources. There are a lot of text sources too like news websites, blogs, e-books, etc. Considering the abundance of unpaired audio and text data for Bangla language, a semi-supervised training method that can exploit both unpaired audio and text is very useful. Proper use of the unpaired data along with existing paired speech corpus can boost the performance of the Bangla ASR system. Different researchers have tried different ways of incorporating this unlabelled, unannotated data for speech recognition. Our approach is similar to the approach used by Karita et al. (2018). We utilize an intermediate representation of speech and text data using a shared encoder network for semi-supervised training of the ASR system. Our contributions in …

Content truncated to 500 words