When the speech reco finally works we'll see...

Is there anything more attention-getting than a software exec gushing about the future when we finally get speech recognition working to perfection? This article in the online Guardian delivers an interview with an unnamed (for obvious reasons) exec at Microsoft who touts virtual secretaries as an application of avatar+speech reco technology. Thanks to Todd Chapin for forwarding this article.

The Guardian reporter is suitable skeptical, which is a nice change from most articles that deal with these sorts of predictions. To get this to work right Microsoft will need to have solved the general AI problem, which is to produce a human-level-or-better intelligence in a machine. If that happens Microsoft won't be wasting its time producing virtual assistants.

When I try to visualize the avatar the exec is talking about, I get an image of the funny Oddcast avatar that I'd written about previously.

Oddcast avatar

This text-to-speech demo by Oddcast is interesting in a couple of ways. Text to speech has improved quite a bit in just the last 3 or 4 years. This demo shows how easy it's become for non specialists to manipulate the sound of a TTS voice in different ways. I must say, some of the accents are less than convincing, but are fun to play with nevertheless. The avatar's responses to the movement of the pointer are quite natural and the timing for returning its gaze to you is well done.

Follow the SitePal link on the upper right side of the page and you'll discover that the company is promoting these avatars as a means of increasing conversion rates on e-commerce web sites. They even provide a high level description of a study. I'm skeptical. I'd really like to see some larger, more-conclusive studies before I'm convinced that an avatar can improve conversion rates. And I'd like an explanation as to why they work.

The demo raises questions about a hypothesis called the The Uncanny Valley phenomenon. To quote the wikipedia definition, the hypothesis states:

As a robot is made more humanlike in its appearance and motion, the emotional response from a human being to the robot will become increasingly positive and empathic, until a point is reached beyond which the response quickly becomes that of strong repulsion. However, as the appearance and motion continue to become less distinguishable from a human being, the emotional response becomes positive once more and approaches human-to-human empathy level.

The Max Headroom videos deliberately exploited the creepiness of a close-but-not-human avatar. The Uncanny Valley is an attractive hypothesis, but there's not a lot of real data to support it. On the other hand, I've seen badly-implemented, unironic Max Headroom-ish trying-too-hard-to-look-real avatars on the websites of major companies and wondered, "what are those people thinking? Have they tested that? That thing is TWITCHING and STARING at me!" The Uncanny Valley was at work. The Oddcast demo is really well done, in part because its not real enough to fall into the Uncanny Valley, but I'm curious about the real business benefit of these avatars.