QASR: QCRI Aljazeera Speech Resource A Large Scale Annotated Arabic Speech Corpus

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2274–2285 August 1–6, 2021. ©2021 Association for Computational Linguistics 2274 QASR: QCRI Aljazeera Speech Resource A Large Scale Annotated Arabic Speech Corpus Hamdy Mubarak,1 Amir Hussein,2 Shammur Absar Chowdhury,1 and Ahmed Ali1 1 Qatar Computing Research Institute, HBKU, Doha, Qatar 2 Kanari AI, California, USA info@arabicspeech.org, https://arabicspeech.org/ Abstract We introduce the largest transcribed Arabic speech corpus, QASR1, collected from the broadcast domain. This multi-dialect speech dataset contains 2, 000 hours of speech sam- pled at 16kHz crawled from Aljazeera news channel. The dataset is released with lightly supervised transcriptions, aligned with the audio segments. Unlike previous datasets, QASR contains linguistically motivated seg- mentation, punctuation, speaker information among others. QASR is suitable for train- ing and evaluating speech recognition sys- tems, acoustics- and/or linguistics- based Ara- bic dialect identification, punctuation restora- tion, speaker identification, speaker linking, and potentially other NLP modules for spoken data. In addition to QASR transcription, we re- lease a dataset of 130M words to aid in design- ing and training a better language model. We show that end-to-end automatic speech recog- nition trained on QASR reports a competi- tive word error rate compared to the previous MGB-2 corpus. We report baseline results for downstream natural language processing tasks such as named entity recognition using speech transcript. We also report the first baseline for Arabic punctuation restoration. We make the corpus available for the research community. 1 Introduction Research on Automatic Speech Recognition (ASR) has attracted a lot of attention in recent years (Chiu et al., 2018; Watanabe et al., 2018). Such success has brought remarkable improvements in reaching human-level performance (Xiong et al., 2016; Saon et al., 2017; Hussein et al., 2021). This has been achieved by the development of large spoken cor- pora: supervised (Panayotov et al., 2015; Ardila et al., 2019); semi-supervised (Bell et al., 2015; Ali 1QASR ����� in Arabic means “Palace”. The acronym stands for: QCRI Aljazeera Speech Resource. et al., 2016); and more recently unsupervised (Valk and Alum¨ae, 2020; Wang et al., 2021) transcrip- tion. This work enables to either reduce Word Error Rate (WER) considerably or extract metadata from speech: dialect-identification (Shon et al., 2020); speaker-identification (Shon et al., 2019); and code- switching (Chowdhury et al., 2020b, 2021). Natural Language Processing (NLP), on the other hand values large amount of textual infor- mation for designing experiments. NLP research for Arabic has achieved a milestone in the last few years in morphological disambiguation, Named En- tity Recognition (NER) and diacritization (Pasha et al., 2014; Abdelali et al., 2016; Mubarak et al., 2019). The NLP stack for Modern Standard Ara- bic (MSA) has reached very high performance in many tasks. With the rise of Dialectal Arabic (DA) content online, more resources and models have been built to study DA textual dialect identification (Abdul-Mageed et al., 2020; Samih et al., 2017). Our objective is to release the first Arabic speech and …

Content truncated to 500 words