Mixture Of Informed Experts For Multilingual Speech Recognition

When trained on related or low-resource languages, multilin-gual speech recognition models often outperform their mono-lingual counterparts. However, these models can suffer fromloss in performance for high resource or unrelated languages.We investigate the use of a mixture-of-experts approach toassign per-language parameters in the model to increase net-work capacity in a structured fashion. We introduce a novelvariant of this approach, ‘informed experts’, which attemptsto tackle inter-task conflicts by eliminating gradients fromother tasks in the these task-specific parameters. We conductexperiments on a real-world task with English, French andfour dialects of Arabic to show the effectiveness of our ap-proach. Our model matches or outperforms the monolingualmodels for almost all languages, with gains of as much as31% relative. Our model also outperforms the baseline mul-tilingual model for all languages, with gains as large as 9%