JHU IWSLT 2022 Dialect Speech Translation System Description

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 319 - 326 May 26-27, 2022 c⃝2022 Association for Computational Linguistics JHU IWSLT 2022 Dialect Speech Translation System Description Jinyi Yang†∗ Amir Hussein†∗ Matthew Wiesner‡ Sanjeev Khudanpur†‡ †Johns Hopkins University ‡ Human Language Technology Center of Excellence {jyang126,ahussei6,wiesner,khudanpur}@jhu.edu Abstract This paper details the Johns Hopkins speech translation (ST) system used in the IWLST2022 dialect speech translation task. Our system uses a cascade of automatic speech recognition (ASR) and machine translation (MT). We use a Conformer model for ASR systems and a Transformer model for machine translation. Surprisingly, we found that while using additional ASR training data resulted in only a negligible change in performance as measured by BLEU or word error rate (WER), aggressive text normalization improved BLEU more significantly. We also describe an approach, similar to back-translation, for improving performance using synthetic dialect source text produced from source sentences in mismatched dialects. 1 Introduction In this paper we describe the JHU dialect speech translation submissions and their development. Di- alects are varieties of a language spoken by a group of people, often in a specific geographic location. In many languages, standard rules of pronunciation, orthography and syntax, but also available data re- sources are drawn from a single dominant dialect. A challenge for all language technologies, includ- ing automatic speech recognition (ASR), machine translation (MT), and speech translation (ST), is how to deal with non-standard dialects for which no formal orthography, grammar, or even data exist. Because many dialects are rarely if ever written, evaluation of ASR and MT on dialect speech is not even particularly well defined. However, there are no such problems evaluating speech translation on dialect speech, which here refers to the task of pro- ducing target language text from source language audio inputs. A focus of both the dialect speech translation task and our system development, is how to lever- age available resources from the standard dialect ∗Equal contribution. to improve performance on non-standard dialects. The dialect translation task focuses specifically on Tunisian Arabic. Arabic and its dialects lie on a dialect contin- uum unified by a single standardized dialect, Mod- ern standard Arabic (MSA) (Badawi et al., 2013). MSA is the primary language of formal and writ- ten communications (e.g. news broadcasts, parlia- ments and religion). However, most native Arabic speakers use local dialects in daily life, which gen- erally lack a standard written form. Certain dialects, such as Algerian, Tunisian, and Moroccan Arabic also have strong Romance, and Berber substrates, and may exhibit a high degree of code-switching, especially with French. Traditionally, speech translation systems have been built by cascading ASR and MT models to form a speech translation chain (Dixon et al., 2011). However, the more recent end-to-end approach (Be- rard et al., 2016; Weiss et al., 2017), which directly translates the source speech to target text, is appeal- ing for this task since since both ASR and MT are ill-defined for unwritten spoken dialects, and there were relatively …

Content truncated to 500 words