Beyond Siri: simplifying commands
Showing off to non-iPhone owning friends has never been easier.
Pick up your phone in the pub, confidently say 'Siri, what's the circumference of the Earth divided by the radius of the Moon?' and barely seconds later, you're the only one there who knows the answer is 23.065.
It's a magical experience, and a great toy.
Compared to what we'll have in a couple of generations of phones, though, it's a Speak & Spell. Best of all, voice is just the start of the natural input revolution.
Imagine a world with no keyboards, no tiny buttons, no tutorials and no manuals. You'll just do what comes naturally, and your phone will adapt, using artificial intelligence (AI) to deduce that you're dictating, or that when you say 'Order take-out', you're going to want Thai that day. Or a million other seamless interactions, combining your camera, location, search, databases, music and more, based on massive databases of information and probabilities and tuned to your personal tastes and past history. It's going to be glorious.
It's also just on the edge of being science fiction at the moment. But how is this kind of natural input unlocking our world in the here-and-now? That's one question Siri can't answer. Fortunately, we can.
Beep. Request. Respond
VOICE CONTROL: If developers can perfect voice control then it will really open up the power of natural input
Like most magic, Siri works by taking an incredibly complex series of actions and hiding them behind a simple flourish.
At its most basic level, pressing Siri's microphone button records a short audio clip of your instruction, which your phone passes to its online servers as a highly compressed audio file. Here, your speech is converted into text and fired back, as a piece of dictation or instruction for your iPhone.
There is, of course, more to it than this – as part of the conversion process, for instance, the server doesn't just send back what it thinks you said, but how confident it is about every word. Artificial intelligence is also required to keep track of the conversation and to maintain context by understanding what you mean by tricky words like 'it' and 'that', or if you were more likely to have said 'we went to see' or 'we went to sea'.
That's the gist, though, and iPhone 4S owners will tell you it often works damn well. At least, it does in the US. One of the few major problems with Siri is that much of the best stuff, like finding a restaurant, has yet to arrive internationally, leaving us with much of the gimmickier stuff.
WOLFRAM ALPHA: Siri gets a lot of its data from Wolfram-Aplha
For now then, the rest of us will have to just imagine asking it to find lunch, having the map plotted directly, and in some cases, even booking a restaurant with nothing more than the word 'Yes'. But give it time, these things will come.
Siri isn't the only tool capable of this, though, and while it is currently the most efficient, the competition works in the same way – just two of them being Nuance's Dragon Go! and the Android-only Iris from Indian startup Dexetra. With Apple's legendary secrecy in full effect, it's often by looking at these that we can see what's going on under the surface, and where Siri is likely to go in future.
An assistant in the cloud
DRAGON GO!: Dragon Go! pre-dates Siri, but does a similar job – with a wider range of search destinations
Knowing how it works, two questions will likely immediately pop into your mind: if all the heavy lifting is happening elsewhere, in the cloud, why do you need an iPhone 4S to use Siri? And why can't it all just work right on the phone?
In truth, the likely answer to the first one is simply 'because Apple wanted a cool selling point for the iPhone 4S'. The original version of Siri was a standalone app that ran on a regular iPhone 4, and on the face of it the latest incarnation isn't doing anything that really requires the more powerful A5 processor. There are future-gazing reasons why Apple might want to restrict it, but precious few non-marketing related ones as it stands now.
What everyone does agree on is the importance of sparing your phone the technical heavy lifting, for two reasons: efficiency and updating.
"The original Iris 1.0 did not use a server, everything was being processed from the phone," explains Narayan Babu, CEO of Dexetra. "Even on powerful phones with dual-core processors, this was inefficient. Natural language processing (NLP) and voice-to-text require real horsepower. When we tried doing serious NLP on Android phones, it almost always crashed. It is also easy to add features seamlessly when processing happens in the cloud, without having to update the actual app."
Those features aren't simply a question of plugging in more information sources for searches, either. The more people who use a tool like Siri, the more powerful it's capable of becoming.
Vlad Sejnoha, chief technology officer of Dragon Go! creator Nuance, one of the most highly regarded companies in the field, told us: "10 years ago, speech recognition systems were trained on a few thousand hours of user speech; today we train on hundreds of thousands. Our systems are [also] adaptive in that they learn about each individual user and get better over time."
To put this into context, speech-to-text tools have been available for many years, but traditionally had to be trained to your voice by having you painstakingly read out long stretches of prose. Modern equivalents still struggle with strong accents, but now failure isn't forever. Over time, their understanding of, for example, a Geordie 'reet' versus a careful Received Pronunciation 'right' can only improve.
Beyond Siri: the next-generation
The soul behind the screen
Giving a computer a name and a voice immediately humanises it. Be honest, have you ever caught yourself thanking Siri, just to be polite? At the very least, do you think of it as a 'him' or a 'her' depending on the voice?
If so, don't worry. It's perfectly normal. "We've seen long conversations ranging from talking about breakups to movies or even philosophy that people have had with Iris," admits Babu.
Siri does have a human face, though – voiceover artist, Weakest Link announcer, and Tap! subscriber Jon Briggs (@jonbriggs on Twitter). How does he feel about his voice becoming our digital butler?
JON BRIGGS: Meet Jon Briggs, the real face of Siri's English (United Kingdom) voice
"I love it," he exclaims. "I love the fact that I have been chosen to be part of people's everyday lives, and especially by a company that creates brilliant technology."
Briggs didn't record his voice specifically for Siri, though – Apple licensed an existing character, 'Daniel', previously used in both Garmin and TomTom sat navs.
The one recording can handle multiple jobs due to being based on individual phonemes (the smallest parts of sound, of which there are 44 in English) and other important parts of the language, rather than specific pre-built statements such as 'turn left'. Combined, these pieces can create more or less any sentence you need.
"We recorded over three weeks – about three hours at a time, then topped up with anything they were missing after it was all analysed," Briggs explains. "The sentences were read as flat as possible, only with intonation where indicated, and no pausing unless there was punctuation. Not as easy as it sounds. Pick a sentence and read it out loud and your pauses won't often fall exclusively where the punctuation is."
As with voice-to-text, it's a technology with a good way to go before it becomes completely reliable, but while Siri may occasionally sound a little sarcastic or irritated, its voices aren't unpleasant to listen to in the long run.
Which of the voices does Briggs himself use? "Which one do you think?!" he answered. We wonder if he ever thanks it…
The next big leap
SHAZAM: Not all natural-input is command based. Shazam can guess almost any music track after just 30 seconds
What all this should demonstrate is that good natural input isn't simply a question of making individual apps that do everything, but creating pieces that can be combined into many forms.
If you want to make an augmented reality system devoted to turn-by-turn walking instead of driving, for instance, you don't have to reinvent the wheel. You know your user will have GPS built into their phone and that you can tap into it, you can give it a professional voice far superior to anything you might whip up yourself, and so on.
When apps can share what they know as easily as they now tap into our Twitter feeds, expect greatness. The catch is that, for now, development is still largely restricted to a bubble. Only Apple can attach data sources and apps to Siri, for example, with everyone else reduced to half-hearted hacks such as using CalDAV calendars to sneak in round the side.
Bouncing between 10 different apps based on what you want to record/look at/scan/find is already frustrating, and is largely self-defeating. Apple, Microsoft, Google… no one company is ever going to create a perfect, all-encompassing natural input system on its own. It's just too big a job.
It's firmly Apple leading the charge, though, and for a glimpse of the future, you can't do better than the iPhone 4S. Siri is at least a top-tier assistant, and no other phone boasts as wide a selection of companion apps, or the same seemingly genuine intelligence.
 |  |
0 comments:
Post a Comment