Development of Automatic Speech Recognition for the Documentation of Cook Islands Māori

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3872–3882 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3872 Development of Automatic Speech Recognition for the Documentation of Cook Islands M¯aori Rolando Coto-Solano1, Sally Akevai Nicholas2, Samiha Datta1, Victoria Quint1, Piripi Wills3, Emma Ngakuravaru Powell4, Liam Koka‘ua3, Syed Tanveer1, Isaac Feldman1 1Dartmouth College, New Hampshire, United States 2Massey University Te Kunenga ki P¯urehuroa, Auckland, Aotearoa New Zealand 3K¯orero Rororuia, Auckland, Aotearoa New Zealand 4University of Otago Te Whare W¯ananga o ¯Ot¯akou, Dunedin, Aotearoa New Zealand rolando.a.coto.solano@dartmouth.edu, s.nicholas@massey.ac.nz emma.powell@otago.ac.nz, {willstutor,liamkokaua}@gmail.com {samiha.datta.23,victoria.r.quint.22,syed.h.tanveer.21,isaac.c.feldman.23}@dartmouth.edu Abstract This paper describes the process of data processing and training of an automatic speech recognition (ASR) system for Cook Islands M¯aori (CIM), an Indigenous language spoken by approximately 22,000 people in the South Pacific. We transcribed four hours of speech from adults and elderly speakers of the language and prepared two experiments. First, we trained three ASR systems: one statistical, Kaldi; and two based on Deep Learning, DeepSpeech and XLSR-Wav2Vec2. Wav2Vec2 tied with Kaldi for lowest character error rate (CER=6±1) and was slightly behind in word error rate (WER=23±2 versus WER=18±2 for Kaldi). This provides evidence that Deep Learning ASR systems are reaching the performance of statistical methods on small datasets, and that they can work effectively with extremely low-resource Indigenous languages like CIM. In the second experiment we used Wav2Vec2 to train models with held-out speakers. While the performance decreased (CER=15±7, WER=46±16), the system still showed considerable learning. We intend to use ASR to accelerate the documentation of CIM, using newly transcribed texts to improve the ASR and also generate teaching and language revitalization materials. The trained model is available under a license based on the Kaitiakitanga License, which provides for non-commercial use while retaining control of the model by the Indigenous community. Keywords: Automatic speech recognition, Cook Islands M¯aori, language documentation, low-resource languages 1. Introduction Automatic Speech Recognition technology could dra- matically accelerate language documentation of In- digenous languages and aid researchers and commu- nity members in the time-consuming task of manual transcription. In this paper we describe the creation of automatic speech recognition (henceforth ASR) for Cook Islands M¯aori, an Indigenous Polynesian lan- guage from the Realm of New Zealand1 Section 1.1 describes the challenges of training ASR for extremely low-resource Indigenous languages, while section 1.2 reviews previous work in natural language processing for Cook Islands M¯aori. Section 2 presents the method- ology for data processing and ASR training using sta- tistical and Deep Learning techniques, and section 3 will describe the training results. Finally, section 4 will describe the future use of this ASR system, as well as li- censing to ensure that the models remain under the con- trol of Indigenous communities and contribute to the development of NLP for other Polynesian languages. 1The Realm of New Zealand includes the Cook Islands, Tokelau and Niue, as well as “New Zealand proper”, or Aotearoa as it is known in Te Reo M¯aori, the M¯aori of New …

Content truncated to 500 words