Speaker 1: Our hypothesis was that combining automatic speech recognition with automatic segmentation of the produced text into subtitles and automatic timecoding of those subtitles would make a useful tool for intralingual subtitling or subtitling for the deaf and hard of hearing. Our subtitlers have quite a lot of experience already with different forms of automatic speech recognition and some of them use speech recognition applications in their daily work already. What was really different in this experiment was the introduction of automatic segmentation and automatic timecoding into the subtitling workflow. For this experiment, we partnered with Lingsoft Language Services for a three-month trial of their subtitling service. And during this trial, our in-house subtitlers would have access to the Lingsoft subtitling service where they would upload their videos and the system would analyze the videos and produce timecoded intralingual subtitles. The subtitler would then download those subtitles in SRT format and insert them into the subtitling software used at YLE to make the necessary corrections and other changes into the text. Time input is always needed in this kind of workflow because even when the speech recognition results are good, even if they're perfect, editing still needs to be made. It needs to be pared down to be condensed so that the viewer can keep up with the text. And here you can see it in the subtitling software. And on the top right, here in red, you can see the reading speed or the reading rate of that subtitle. And that number needs to be below 12 to be up to our specifications. So even if there is no recognition mistakes, as there is none on this screen currently, it needs to be edited in order to fit our criteria of good subtitles. So what did we learn? We learned that this kind of tool works really well with some program types and less well with others. Typically it works well with scripted speech and less well with spontaneous speech. We learned that using this tool can make a subtitler work faster or slower, depending not just on the program type, but on individual preferences and ways of working. We also learned that you really have to be careful when proofreading text produced by automatic speech recognition, because some of the small recognition mistakes can be really easy to miss. Overall the feedback was really positive from our subtitlers. They see a lot of potential in this kind of tool. And everyone who participated in this trial would use such a tool in their daily work if it was available.