Multilingual Modulation by Neural Language Codes

Опубликовано 9 октября 2018, 17:39

Multilingual Speech Recognition is a very costly AI problem, as each language and even different accents require their own acoustic models to obtain optimal recognition performance. Even by using the same phone symbols across languages, each language and even accents impose their own colorings or "twangs", a shift in the acoustic realization of sounds. In this talk, I will outline an approach that uses a large multilingual neural network that is modulated by language codes. These codes are generated by an ancillary network that learns to code useful differences between the "twangs" or human languages. This network architecture allows the quick adaptation to languages.

See more at microsoft.com/en-us/research/v...