Adventures in Audio Data Collection

Ever wonder how Rosetta Stone assesses your pronunciation? In a nutshell, your speech is compared to native speakers’ pronunciation to determine how accurate your response was. How do we know what native speakers sound like? We collect lots of speech data from real people to diseartlearn the native sounds of a language. Without native data, we wouldn’t have anything to compare your speech to. Gathering this native speech data is necessary to develop the speech recognizer for all of our languages, which can be challenging in some cases – for example, Irish. Lots of Americans have Irish heritage and connections to Ireland, but it’s rare to find a native speaker in the US, let alone a large group of native speakers who are willing to record an hour’s worth of audio data within the span of a few weeks.

The preparation for the launch of the Irish product happened back when all of our audio data had to be collected in-person using a locally installed data collection tool. In those days, we had no way to collect speech data over the web as we do today. This meant that I needed to find  Irish speakers somewhere in the Denver/Boulder metro area. Considering all of these factors, it slowly dawned on me: this was going to be impossible. Almost as a joke, I said to Bryan Pellom (VP, Speech Development), “Let’s just go to Ireland”. That was in August 2007; by November (at which point I had located exactly one Irish speaker living in Denver) we were landing at the Shannon Airport in Ireland.

Where in Ireland do you go when you need to find a high concentration of Munster Irish speakers who are enthusiastic about their language and will do just about anything to promote it? It turns out there’s a good answer to that question:  Dingle! And once you’re in Dingle, look no further than the Díseart Institute of Education and Celtic Culture to find your Irish language enthusiasts.  The Díseart Institute, founded in 1998, now occupies the former Presentation Sisters’ convent in a magnificent neo-gothic building which was home to a community of sisters as early as 1829 right in the heart of Dingle.

data collection headquarters at the diseartI never would have guessed it, but it turns out that former convents make great places to record audio data! There are lots of individual rooms (former dormitories) and they tend to be quiet, peaceful places. So, with plenty of room to spread out and a laptop and headset each, Bryan and I began the task of recording people all day long for five straight days. We set up one of the dormitory rooms as our headquarters and spent the next week camped out in the convent.

The fact that Dingle is a relatively small town definitely played to our advantage. It seemed like everyone in town knew that Rosetta Stone was looking for Irish speakers after the first day of recording and wanted to get involved. I’m not sure if we would have been as successful or experienced the same level of excitement from the community if we had chosen to record in a larger town. It was also amazing to get that kind of glimpse into Dingle life; by the end of the week, I was running into our participants all over town and I felt just like an honorary Dingle resident. Among those we recorded were Díseart employees and members of their Board of Directors, numerous local farmers and teachers, a master crystal cutter, several local musicians and a canon.  Yes, indeed – when you speak to the Rosetta Stone Irish product, your pronunciation is being compared to that of a canon!

On the whole, the recording process went amazingly well once we hit our stride. That’s not to say that there were no challenges along the way! We hadn’t predicted that many of our participants would have grown up speaking Irish at home, but not necessarily reading it (and Irish is not an easily-read language!). For these participants, the process was slower, but they were able to use the audio from our collection tool to memorize and repeat each phrase, one at a time. Other participants were not comfortable using our laptops and were worried that they wouldn’t be able to participate. We solved that problem by controlling the computer mouse to do all the navigation for them, while still allowing them to contribute their audio to the project.

meagan sills phil easleyThese days, most of our audio data comes in through the web. We’re currently recording everything from British English to Pashto and Swahili.  While the web allows us to be more efficient in many ways, I’m so grateful for the on-site collection experience we had in Dingle. I still feel connected to the Rosetta Stone Irish product and the community in Dingle, which is a place that you fall in love with right away and never forget.  If you ever find yourself in County Kerry, it’s worth your time to head to the Dingle peninsula and visit the Díseart gift shop or take a tour of the Harry Clarke stained glass windows in the convent. Bring your Rosetta Stone Irish product box and ask for their autograph, because those folks contributed in so many ways to this project and would be thrilled to converse with you in Irish.

Find more posts about:

Meagan Sills

Language Resource Manager Meagan Sills is responsible for collecting native and non-native speech data which is used to develop Rosetta Stone’s speech recognition engine. She joined the company in 2006 and works from the Boulder, Colorado office. Meagan has an M.A. in French Literature from Ohio University.
blog comments powered by Disqus