YouTube now Adding Close Captioning Automatically
This is an excellent step in the right direction to add more accessibility to the second most popular search engine on the planet. Deaf and hearing-impaired gamers will now be able to begin looking up cheat codes for their favorite video games just like everyone else!
"Google (NSDQ:GOOG) has offered user-generated captioned videos for three years. What makes "auto-caps" different is that Google will now use the speech-recognition algorithms employed in Google Voice to automatically generate captions for all videos...
In addition, Google announced automatic caption timing, or auto-timing, which eases the process of creating captions manually. Video creators build and upload a text file with all the words in the video. Google's automatic speech recognition (ASR) technology discerns when the words are spoken and creates the captions. Google hope the technology will bring even more captioned videos to the site since it significantly reduces the cost in time and resources necessary to create professional caption tracks."
You can see the entire article on CNN.
TrackBack URI for this entry
Subscribe to this comment's feed
Unfortunately, this is one of those things that sounds a lot better on paper than it does in practice.
I tried it on a video about adaptive sports on UCLA's channel, and the phrase "we're doing rock climbing as well as table tennis," pronounced in a perfectly neutral Hollywood-style accent, was transcribed as "for doing what I mean as well as Democrats." Also, "wheelchair basketball" became "mister bass one."
Perhaps as the technology gets better this will improve, but for right now, it's like trying to read subtitles on a bad bootleg DVD...
The variance in accuracy encountered when having voicemails automatically transcribed is huge, so assuming this uses the same acoustic model, we can't hope for better results. At least not right away.
Fortunately, if anybody can pull it off, it's Google. As I understand it, the speech corpus they've assembled is beyond massive. There is simply no way that even half of it has been incorporated into their acoustic model yet, so we should be able to count on steady improvement in recognition accuracy.
