Code-Switching

While most work in automatic speech recognition focuses on individual languages, it is common for multilingual speakers to switch between languages within conversations. Even for languages where this is common, available datasets often do not reflect this phenomenon. In fact, speakers are often requested not to code-switch during data collection.

Building speech recognition systems that handle code-switching must deal with the lack of training data. Our focus has been on building models without any available code-switched data. Instead we leverage multiple out-of-domain monolingual sources to accomplish the task.

Related Publications

• “Using Heterogeneity in Semi-Supervised Transcription Hypotheses to Improve Code-Switched Speech Recognition”, Andrew Slottje, Shannon Wotherspoon, William Hartmann, Matthew Snover, and Owen Kimball, arXiv preprint arXiv:2106.07699, 2021. [arxiv] [bib] [post]