Posts Tagged ‘siri’

BBC News – Say what? iPhone has problems with Scots accents

The new iPhone 4S came out a few weeks ago to a relatively muted reception, primarily because it wasn’t the massive technological leap ahead that most people has expected (or hoped for). Nevertheless, one much vaunted tool was the new software ‘Siri‘, a natural language processing programme which is able to do everything from noting your reminders to showing you where the nearest cash point is to searching the internet, simply through you talking directly to it. Now, natural language processing is kind of viewed as the holy grail of search engines, primarily because when we want to find something, we don’t think in terms of key-words but rather coherent sentences. Ask Jeeves was probably one of the more successful attempts at integrating NLP processes as part of its search optimisation, but it wasn’t able to knock Google off its perch as the de facto search engine. This is because one of the major difficulties in developing intelligent NLP software in search engines is that it has to be able to deal with a variety of inputs. For instances, if I wanted to find out how to make homebew, it could be phrased in a number of different ways ‘How to make homebrew’, ‘How do I make homebrew?’, ‘What do I need to start homebrewing’, and so on. Google, on the other hand, doesn’t use NLP but keyword searches. So, for the previous example, Google would trawl its indexed webpages looking for instances of ‘homebrew’, ‘starting’ and so on, and pages which had more of these keywords would appear higher up the search rank. Now, this works fine for simple searches, but you enter a question into Google and it starts to struggle, outputting quit useless pages which aren’t really relevant to your query (I can’t think of any examples off the top of my head, but I know it’s happened a bunch of times to me…).

When we get to voice activated NLP, we get into even more difficulty. Not only does the software have to deal with different forms of input in terms of lexical items and grammatical structure, but it also has to deal with variability of accent and dialect. For example, it has to be able to identify the fact that /bʌtəɹ/ and /bʌʔəɹ/ are the same word (butter), even though it’s pronounced slightly differently. Again, that’s only a simple example, but when we think of the variability of phonological structure between different varieties, the software has to be powerful enough to store these different varieties and recognise them when they’re input into the system. Take into account differences of age, sex, class and so on, what Siri can do gets ever more impressive. That is, until it tries to deal with Scots.

Now, obviously there are a number of stereotypes about Scottish speakers and how difficult they are to understand (when I play clips of Scottish English speakers to my students, there is often a stunned silence in the room as they try to decipher whether the speaker is asking for the time or threatening to chib someone…). And it seems that Siri is similarly affected (for the most part) in trying to deal with Scottish accents (BBC News – Say what? iPhone has problems with Scots accents). I can’t say that this is entirely unsurprising given the amount of trouble I have with voice recognition technology. You know, the kind that’s used in telephone banking, cinema booking lines and car insurance call centres… Quite why the technology hasn’t caught up with understanding Scots is a bit of a mystery to me, especially since it should technically be the same kinds of processes of programming the software as would be followed for Souther Standard British English or General American. I mean, I get that Scotland is a small country in the grand scheme of things, but it’s kind of difficult to ignore a potential market of 6 million speakers.

But I suppose that so long as we’re not lumbered with voice activated lifts, we’ll be alright…