Michiel Bacchiani / Google
■ Session Overview
When neural networks re-gained popularity in speech recognition about 10 years ago they were mainly used for the acoustic model of the system (the model that relates the audio with phonetic units). To obtain a complete recognition system, those models would be combined with a language and pronunciation model. Due to ongoing research, recent years have shown that speech recognition systems can be built that is a singular neural network that encompasses the entire speech to text system, the so-called end-to-end systems. These models are of interest as they are compact, accurate due to their joint optimization and easy to build as there is very little need for manual design. On the other hand, in contrast to the previous systems, they have given rise to a number of research problems related to control and online operation of such models. This talk will describe some of the research Google has done to address such issues.
■ Official Site #linedevday
linedevday.linecorp.com/2020/...
■ Category
AI