Automatic Speech Recognition for Uyghur through Multilingual Acoustic Modeling

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6444–6449 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6444 Automatic Speech Recognition for Uyghur through Multilingual Acoustic Modeling Ayimunishagu Abulimiti, Tanja Schultz Cognitive Systems Lab, University of Bremen, Germany {ay.abulimiti, tanja.schultz}@uni-bremen.de Abstract Low-resource languages suffer from lower performance of Automatic Speech Recognition (ASR) system due to the lack of data. As a common approach, multilingual training has been applied to achieve more context coverage and has shown better performance over the monolingual training (Heigold et al., 2013). However, the difference between the donor language and the target language may distort the acoustic model trained with multilingual data, especially when much larger amount of data from donor languages is used for training the models of low-resource language. This paper presents our effort towards improving the performance of ASR system for the under- resourced Uyghur language with multilingual acoustic training. For the developing of multilingual speech recognition system for Uyghur, we used Turkish as donor language, which we selected from GlobalPhone corpus as the most similar language to Uyghur. By generating subsets of Uyghur training data, we explored the performance of multilingual speech recognition systems trained with different sizes of Uyghur and Turkish data. The best speech recognition system for Uyghur is achieved by multilingual training using all Uyghur data (10 hours) and 17 hours of Turkish data and the WER is 19.17%, which corresponds to 4.95% relative improvement over monolingual training. Keywords: ASR, Low- resource, Multilingual training, Agglutinative languages, GlobalPhone 1. Introduction Speech recognition technology has gained dramatic im- provement recently and has shown promising results on many tasks (Graves et al., 2006; Hinton et al., 2012; Chiu et al., 2018). However, data sparsity is still an issue for train- ing more reliable acoustic and language models. Compared to resource-rich languages, low-resource languages suffer from higher Word Error Rate (WER) in speech recognition tasks. It has been long established that acoustic models trained across multiple languages can partly compensate this resources gap (Schultz and Waibel, 2001). In particu- lar, data from resource-rich languages, which are similar to the target low-resource language, might be used in multi- lingual training with the aim to cover more contextual vari- ations, thereby improving the performance of Automatic Speech Recognition (ASR) system for low-resource lan- guages (Caruana, 1997; Heigold et al., 2013; Ghahremani et al., 2017; Huang et al., 2013; Sahraeian and Van Com- pernolle, 2016). However, the impurity of training data that comes from different sources (donor languages and target language), may hurt the acoustic model of target language and even the donor language (Vu and Schultz, 2013; Lin et al., 2009). Therefore, finding proper balance between the amount of data from target and donor language is one issue in developing multilingual ASR for low-resource language. Uyghur language is an under-resourced language with about 11 million speakers, who are mainly located in western China and Central Asia. Uyghur belongs to the Turkic fam- ily. It is …

Content truncated to 500 words