Apple's Siri team has published a trio of posts on Apple's Machine Learning Journal. The posts detail how the team improves Siri, covering topics like introducing new Siri languages, how Siri displays data like times, dates, and addresses, and the evolution of Siri's voice.
From Vol. 1, Issue 3, Inverse Text Normalization as a Labeling Problem:
Siri displays entities like dates, times, addresses and currency amounts in a nicely formatted way. This is the result of the application of a process called inverse text normalization (ITN) to the output of a core speech recognition component. To understand the important role ITN plays, consider that, without it, Siri would display "October twenty third twenty sixteen" instead of "October 23, 2016". In this work, we show that ITN can be formulated as a labeling problem, allowing for the application of a statistical model that is relatively simple, compact, fast to train, and fast to apply.
Apple introduced the Machine Learning Journal just over a month ago as a way to more publicly share its research and efforts in the area of machine learning. The first article in the journal centered around Apple's use of synthetic images to train neural nets.
