Towards the automatic processing of Yongning Na (Sino-Tibetan): developing a 'light' acoustic model of the target language and testing 'heavyweight' models from five national languages

Thi-Ngoc-Diep Do (1) , Alexis Michaud (2) , Eric Castelli (1)

1 Speech Communication
2 Speech Communication

Automatic speech processing technologies hold great potential to facilitate the urgent task of documenting the world's languages. The present research aims to explore the application of speech recognition tools to a little-documented language, with a view to facilitating processes of annotation, transcription and linguistic analysis. The target language is Yongning Na (a.k.a. Mosuo), an unwritten Sino-Tibetan language with less than 50,000 speakers. An acoustic model of Na was built using CMU Sphinx. In addition to this 'light' model, trained on a small data set (only 4 hours of speech from 1 speaker), 'heavyweight' models from five national languages (English, French, Chinese, Vietnamese and Khmer) were also applied to the same data. Preliminary results are reported, and perspectives for the long road ahead are outlined.

Mots clés: Acoustic models; automatic speech recognition (ASR); multilingual modelling; under-resourced languages; endangered languages; Yongning Na; Naish languages; language portability; statistical language modeling; crosslingual acoustic modelling and adaptation

Domaines: Linguistique; Traitement du signal et de l'image [eess.SP]; Traitement du signal et de l'image [eess.SP]

Thi-Ngoc-Diep Do, Alexis Michaud, Eric Castelli. Towards the automatic processing of Yongning Na (Sino-Tibetan): developing a 'light' acoustic model of the target language and testing 'heavyweight' models from five national languages. 4th International Workshop on Spoken Language Technologies for Under-resourced Languages (SLTU 2014), May 2014, St Petersburg, Russia. pp.153-160. ⟨halshs-00980431v2⟩