Automatic Transcription Challenges for Inuktitut, a Low-Resource Polysynthetic Language

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 2521–2527 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 2521 Automatic Transcription Challenges for Inuktitut, a Low-Resource Polysynthetic Language Vishwa Gupta, Gilles Boulianne Centre de recherche informatique de Montr´eal (CRIM) {vishwa.gupta, gilles.boulianne}@crim.ca Abstract We introduce the first attempt at automatic speech recognition (ASR) in Inuktitut, as a representative for polysynthetic, low-resource languages, like many of the 900 Indigenous languages spoken in the Americas. As most previous work on Inuktitut, we use texts from parliament proceedings, but in addition we have access to 23 hours of transcribed oral stories. With this corpus, we show that Inuktitut displays a much higher degree of polysynthesis than other agglutinative languages usually considered in ASR, such as Finnish or Turkish. Even with a vocabulary of 1.3 million words derived from proceedings and stories, held-out stories have more than 60% of words out-of-vocabulary. We train bi-directional LSTM acoustic models, then investigate word and subword units, morphemes and syllables, and a deep neural network that finds word boundaries in subword sequences. We show that acoustic decoding using syllables marked with word boundary markers results in the lowest word error rate. Keywords: DNN, deep neural networks, automated transcription of Inuktitut, polysynthetic language, low resource language 1. Introduction Inuktitut is one of the 60 Indigenous languages from 12 distinct language families currently spoken in Canada, and approximately 36,000 people declare Inuktitut as their mother tongue1. There is a growing interest shown by In- digenous communities, and federal and provincial govern- ments, in the development of language technology (Littell et al., 2018), as it could help revitalize historical recorded archives, provide language course materials, and create apps to promote language dissemination among the young. Inuktitut is described as a highly polysynthetic language: its words are composed of many morphemes, such that single words can express what usually requires a whole sentence in other languages. Polysynthetic languages are often termed agglutinative when their morphemes have clear boundaries and thus are easily segmentable (Kla- vans, 2018). It is not entirely the case in Inuktitut, where morphemes combine in rich and complex ways that af- fect their pronunciation, giving rise to a complicated map- ping between surface segmentation and underlying mor- phemes (Micher, 2017). Previous work in speech recognition have investigated sev- eral polysynthetic agglutinative, resource-rich languages such as Finnish, Estonian, Turkish, Korean, or Hungar- ian (Mihajlik et al., 2007) (Kurimo et al., 2007) (Erdoˇgan et al., 2005) (Kwon and Hwang, 1999). Speech recogni- tion systems usually rely on a large enough word lexicon to cover most words in an unseen text, as each occurrence of an out-of-vocabulary word will cause at least one recogni- tion error. Polysynthetic languages challenge this approach since they easily generate very large number of distinct words through combination of morphemes. Many studies resort to subword units in order to increase the coverage with a reasonable size subword lexicon. Sev- eral possible subword units have been studied, includ- ing syllables …

Content truncated to 500 words