2021/09/21 by Saurav Jha, Jha, Saurav · 1 citation
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Speech Recognition and Synthesis #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2110.10261
openalex publication_date 2021/09/21 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Automatic Speech Recognition (ASR) systems have been gaining popularity in\nthe recent years for their widespread usage in smart phones and speakers.\nBuilding ASR systems for task-specific scenarios is subject to the availability\nof utterances that adhere to the style of the task as well as the language in\nquestion. In our work, we target such a scenario wherein task-specific text\ndata is available in a language that is different from the target language in\nwhich an ASR Language Model (LM) is expected. We use Neural Machine Translation\n(NMT) as an intermediate step to first obtain translations of the task-specific\ntext data. We then train LMs on the 1-best and N-best translations and study\nways to improve on such a baseline LM. We develop a procedure to derive word\nconfusion networks from NMT beam search graphs and evaluate LMs trained on\nthese confusion networks. With experiments on the WMT20 chat translation task\ndataset, we demonstrate that NMT confusion networks can help to reduce the\nperplexity of both n-gram and recurrent neural network LMs compared to those\ntrained only on N-best translations.\n