1

We present the first unified study of the efficiency of self-attention-based Transformer variants spanning text, speech and vision. We …

Automatic speech recognition research focuses on training and evaluating on static datasets. Yet, as speech models are increasingly …

Recent visuolinguistic pre-trained models show promising progress on various end tasks such as image retrieval and video captioning. …

For the majority of the machine learning community, the expensive nature of collecting high-quality human-annotated data and the …

We present a reduce-and-reconstruct paradigm that compresses the output vocabulary of the ASR system using linguistically meaningful reductions and then reconstructs the original vocabulary using a standalone FST module.

We propose a new transfer learning strategy by pretraining using large amounts of speech in the high-resource language with its text transliterated to the target low-resource language, obtaining relative WER reductions of up to 8.2% on 1 hr of target speech compared to existing transfer-learning approaches.

We release ~600 hrs of transcribed speech data in 7 Indian languages, including 2 code-switched language pairs. We also provide baseline Kaldi and ESPNET recipes for both the subtasks with 30.73% and 32.45% WERs on the multilingual and code-switching test sets respectively.