If you’ve ever used voice search on your Fire TV and noticed results that are completely unrelated to your request, that’s actually being done intentionally. Amazon has revealed a recent improvement to the Fire TV’s voice search, called Phonetically Blended Results (PBR), that deliberately mix in results for similar sounding queries, even though it may have nothing to do with the words the Fire TV thinks you said.
The way PBR improves search results by, technically, first making them a little worse, starts with the admission that the translation from a voice query to text isn’t perfect. Most Fire TV owners have said something into the microphone of their voice remote, only to see something similar sounding but completely different appear on the screen.
An example Amazon gives where PBR helps is if someone says “Encanto” with loud kids in the room and it is interpreted as “Encounter” by the Fire TV voice remote. Prior to the launch of PBR on the Fire TV, only results matching “Encounter” would be displayed. Now, PBR mixes in a few similar-sounding movies and TV shows, like “Encanto,” into the list of results shown on the screen, just in case the voice translation to text wasn’t perfect.
Amazon says “about 80% of the 20 million or so unique search terms that Fire TV deals with are augmented by PBR.” The system learns which similar-sounding words should be mixed together in search results by paying attention to how people interact with voice search. If someone makes a voice request for one thing and then immediately makes a second voice request for something different that leads to a selection, Amazon’s system takes note that those two requests are probably related. After the same pattern of back-to-back requests happens enough times, the search results for both requests will become mixed together so that the results from the second request are already present when making the first request.
The actual system in play is much more complex than the above example. Each search-query mapping is given a PBR confidence score that is used to determine if the results of other search terms should be mixed in and how many of them. For example, Amazon says that a search for “Enchanted” used to result in a selection only 60% of the time because customers were looking for “Encanto” 20% of the time and “Disenchanted” 5% of the time. Now, depending on many factors, those alternative results can be mixed in with the request for “Enchanted” to improve the likelihood that a selection will be made with just a single search.