More Tales from Startup Alley: Aurality Kinda Reads the News

I’ve always liked the idea of having an app that would read the New York Times to me as I was unwinding on the 5:45 back to Jersey. I though automated news software made sense for incredibly busy Type A personalities who were racing around doing other things and busy news-addicted bloggers. So I was inclined to be very excited by Aurality, a news reading app for the iPhone that I came across at TechCrunch Disrupt NY.

Awhile back I experimented with my own text-to-speech software using APIs from Twilio and Voxeo. In fact, I came up with a bare-bones headline reading app, which I was irrationally proud of.

I learned the hard way that any news reading approach would have to both understand the context of the article–is it a sports story or a technology piece, etc.– and master something called phonemes, more on that in a bit.

I downloaded the free Aurality app onto my iPad and gave it a few articles from the New York Times business and sports sections. The actual selection of the article and navigation in Aurality is smooth enough. Their backend software will do some processing on the text before it reads it–this can take a few seconds so you have to be a little patient.

Presumably, Aurality is marking up the text with tags from VoXML (aka, Voice XML), which is the open standard for speech processing. While anybody can do a straight dump of the text into a VoXML page, it requires careful parsing of the story to get resulting speech to sound better than Esperanto.

Here’s what I discovered about Aurality’s news reading.

It doesn’t understand the essential fact many news stories have a dateline–i.e., a city or other location. Aurality read a story about the Greek financial crisis by first articulating the words “London dash.” Savvier software would have said “Dateline London”.

Aurality has the annoying habit of reading quotes by saying “open double quotes” and then terminating the sentence with a coldly robotic “close double quotes”. It also hasn’t figured than an em-dash doesn’t needed to be literally spoken back–it just requires a pause.

And don’t ask about sport scores and team names. I winced as it reported a “105 dash 93” victory for the Miami Heat.

It gets worse. It seems to ignore words that have hyperlinks–that’s just a basic parsing error on Aurality’s part.

Then there’s the whole issue of phonemes, which is the actual pronunciation and emphasis of a word–the markup you see in a dictionary right before the derivation and definition. Is it a hard ‘c’ or a soft ‘c’? Long e, like in repeat or a short e, as in fed? Aurality as a non-native speaker needs this kind of information.

VoXML does have the special phoneme tags to mark up the text appropriately so that speech engine would do a credible job simulating English. But that would require Aurality do more sophisticated natural language parsing. Which they are not doing at this stage

This is a Google-size problem, and so I’m not completely surprised Aurality didn’t get it right. However, for given contexts, say business or sports or technology, Aurality should and could perform better than it currently does.