Live Rewind is surprisingly accurate
I was skeptical, but Live Rewind does a surprisingly decent job of understanding and transcribing what it hears.
I started by playing YouTube videos of restaurant conversations for people learning English as a foreign language: slow, clear, with largely simple vocabulary. I sat close to the iPad playing the video. An easy task to begin with, and Live Rewind coped easily. The transcripts were virtually perfect, not to mention beautifully formatted with line breaks to denote separate short speeches.
I stepped up the difficulty by looking up some menus online (I selected this fancy restaurant) and acting the part of the waiter myself. I selected esoteric dishes I hadn’t heard of, spoke quickly and was careless with my enunciation, chucked in some ums and ers, varied volume and occasionally mumbled a little. And the feature produced a mostly accurate transcript that would have been extremely useful to a diner who had missed an option on the list.
This time the formatting was more of a challenge–Live Rewind couldn’t keep track of separate sentences, so the whole thing was jumbled together with a sometimes misleading number of commas–and occasionally I had to decipher a soundalike guess when the feature didn’t recognise a term. (“IV” instead of Ivy, “7 and Y” instead of the British rivers Severn and Wye, for example.) But the information was mostly there.
The most difficult task was when I used Live Rewind (after obtaining consent!) to recall snatches of random conversations with my family. This was more mixed. It caught the gist a lot of the time. But there were more soundalike guesses than before, and some sentences were missed entirely. This mostly related to distance, I think, and I’ll come back to that later.
So Live Recall is by no means perfect. But considering the difficulty of the task–handling variations in accent and diction, mumbling and muffled speech, irregular grammar, slang, in-jokes, proper nouns, brand names, technical or foreign terms–it does a very solid job. My verdict is that the accuracy is high enough to be useful as backup if you missed something, but not yet accurate enough to depend on as an alternative to paying attention.
