Very few annotated recordings of children reading exist, so this company had to build its own dataset, and annotating it in-house couldn't keep up.
This company was building speech recognition to catch mispronunciations when kids read aloud in its app. There are very few annotated recordings of children reading, so they had to build their own dataset.
They started annotating the recordings in-house, and it was clear pretty quickly that the turnaround wasn't sustainable for the amount of data they needed.
We took on the annotation with our mix of AI and people, set up around their recordings and delivered in the format their team needed.
That gave them more annotators than they had in-house, so their team could keep building the dataset their speech recognition needed.
This was one of our earlier projects, from when a lot of our work was building training data for machine learning teams. It's a good example of the hard version of that work: data that doesn't exist yet, labeled carefully enough to train on.
If you need data that doesn't exist yet and your team can't label it fast enough, this is the kind of work we did here. Talk to us about it.
