SRI-B’s Systems for IWSLT 2023 Dialectal and Low-resource Track: Marathi-Hindi Speech Translation

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 449–454 July 13-14, 2023 c⃝2023 Association for Computational Linguistics SRI-B’s systems for IWSLT 2023 Dialectal and Low-resource track: Marathi-Hindi Speech Translation Balaji Radhakrishnan, Saurabh Agrawal, Raj Prakash Gohil, Kiran Praveen, Advait Vinay Dhopeshwarkar, Abhishek Pandey Samsung R&D Institute,Bangalore {balaji.r, saurabh.a, raj.gohil, k.praveen.t, a.dhopeshwar, abhi3.pandey}@samsung.com Abstract This paper describes the speech translation systems SRI-B developed for the IWSLT 2023 Evaluation Campaign Dialectal and Low- resource track: Marathi-Hindi Speech Trans- lation. We propose systems for both the constrained (systems are trained only on the datasets provided by the organizers) and the un- constrained conditions (systems can be trained with any resource). For both the conditions, we build end-to-end speech translation networks comprising of a conformer encoder and a trans- former decoder. Under both the conditions, we leverage Marathi Automatic Speech Recogni- tion (ASR) data to pre-train the encoder and subsequently train the entire model on the speech translation data. Our results demon- strate that pre-training the encoder with ASR data is a key step in significantly improving the speech translation performance. We also show that conformer encoders are inherently superior to its transformer counterparts for speech trans- lation tasks. Our primary submissions achieved a BLEU% score of 31.2 on the constrained con- dition and 32.4 on the unconstrained condition. We secured the top position in the constrained condition and second position in the uncon- strained condition. 1 Introduction Speech translation (ST) is the task of automatically translating a speech signal in a given language into text in another language. While rapid strides have been made in speech translation in recent times, this progress has been restricted to a small number of high resource languages. This progress excludes sizable sections of people who speak languages that have very little speech data available. So, for these speech systems to be beneficial and impactful in the real world, they have to be developed and shown to work on low-resource languages as well. In order to mitigate these issues and encourage research on low-resource languages, IWSLT pro- pose a dialectal and low-resource speech translation track (Agarwal et al., 2023) as a part of their 2023 shared tasks evaluation campaign. While this track includes various low resource languages, we focus our efforts on the Marathi-Hindi language pair. The goal of this task is to translate Marathi speech to it’s corresponding Hindi text. Marathi and Hindi are both Indo-Aryan languages used in India. Even though there were 83 million people across India speaking Marathi as per the 2011 census of India, it lacks sufficient speech data to support modern speech translation systems. This paper discusses our work and submissions on the Marathi-Hindi low-resource speech trans- lation task. Our experiments in this paper focus only on end-to-end architectures. We begin our experiments with a simple end-to-end Transformer and build on this approach with the following key contributions that significantly better our final per- formance: • Encoder pre-training with Marathi ASR data. • Replacing the Transformer encoder …

Content truncated to 500 words