The Alexa Custom Assistant: what we know, and what we don’t know

Two weeks ago, Amazon announced an Alexa Custom Assistant (ACA). I quickly blogged about it and posted on linkedin saying I thought it was a “big deal” and asked if others agreed. Almost half of the respondents agreed, a few disagreed, and a lot said something along the lines of “it depends”. Of course, “it depends” is the accurate answer, so let us look at what we know and what we don’t know and some of my impressions.

1) ACA makes it easier to build a custom assistant. This is true! The Ned Curic blog repeats the point over and over that building a custom assistant isn’t easy and Amazon knows how to do it. But there are probably a dozen or more other companies that can do it too. The question about quality is always raised when services are involved. Will Amazon spend the time and money they do for Alexa to build a custom assistant?

2) What is ACA? The Alexa Custom Assistant looks like Amazon is just taking 3 existing capabilities and combining them: 1) Alexa skills already exist 2) Amazon can build custom wake words, and 3) Amazon can build custom synthesis engines. Custom wake words and synthesis does not appear to be automated so a lot of this is a service. It feels like it may be a quickly cobbled together solution to find more success in the automotive department. My guess is this a way to make fast headway and they will quickly get stronger and more capable.

3) How about a custom domain? To build a custom assistant a custom domain recognizer is needed that represents the queries for a specific type of function. Alexa covers lot of domains very well, and a lot of others not so well. For example, when (a division of Sensory) tested the Sensory custom domain for a microwave against Alexa’s microwave (which should have been a custom domain but did not appear to be), Sensory’s performance was far superior. It appears that the ACA is relying on its SKILLS for customers to develop custom domains. I am not impressed with Alexa Skills. Although many have developed Skills, they are not well utilized, and they perform very poorly. Simple things like adjusting volumes or answering questions within a Skill are flawed. When my family plays Jeopardy about 15% of the question we answer correctly are counted as wrong because Alexa misunderstands us (and we have native born US accents!). Alexa randomly turns off when we play, and the interaction feels like it requires very scripted responses and not natural language.

4) Cost? I didn’t read anything about cost. Amazon will certainly bring you into their cloud and of course they will have cloud-based fees for AWS. My guess is that without a tool like VoiceHub they will need to charge service fees as well.

5) Data and Privacy. According to Amazon “Further, because the Alexa Custom Assistant is built from the ground up with privacy in mind and using Alexa technology, Amazon manages the data…” Wait a second, Amazon manages the data for you…they get the data, so how is this private? I think they are saying they apply the same rigorous standards that they protect Alexa data with (which of course has led to lots of data leaks already). But on top of that the ACA customer is giving up their data to Amazon. At least Bezos is honest in his quote “Your margin is my opportunity”.

6) Targeted market and competitors. I thought it was funny when Sensory’s VoiceHub announcement was turned into a headline: Sensory taking on Amazon and Google. That wasn’t our intent. Actually, we’d LOVE to be the embedded arm for ACA and I bet we can build custom wake words faster and better than Amazon. Also, our VoiceHub could build grammars that are more effective than Alexa skills. Nevertheless, Sensory is not a general-purpose assistant provider and today we have no cloud-based solutions. I believe Amazon is squarely targeting Cerence and Soundhound who the ACA will directly compete against, and the fact that ACA appear to be in their automotive group with Fiat as a first customer seems to support this targeting.

I still think the Alexa Custom Assistant is a big deal. It represents the explosion of the custom or domain specific assistant and shows how broad reaching Amazon wants to be. I think there will be a big separation now based on privacy…there will be the Cloud Custom Assistants that run on AWS or Azure, etc., the Cloud Custom Assistants that run on private clouds, and the embedded Custom Assistants that run on device and pose no privacy risks whatsoever.