Results of testing subtitle workflow - experiments with universalsubtitles.org
Compare the videos below.
1. Original (unsubtitled) full length video, submitted by Dan Buzzo on Fri November 05, 2010
2. Trimmed (first few minutes) automatically subtitled using google/youtube captioning. Low accuracy, generic voice profile, useless mostly hilarious subtitles.
3. Original video, first few minutes are subtitled using our toolset and deployed using universalsubtitles.org.
(click image to launch video)
I produced this by
1) downloading the original video from Vimeo,
2) using quicktime to save a version of the video in the mp4 container,
3) using the commandline tool 'afconvert' on the mp4 version of the video
4) processing the audio with uweAudioSegmentation,
5) using uweRemoteTranscription and uweScribeBatchTranscription to produce a transcript,
6) converting this to SRT using a quick and dirty converter (yet to be comitted to github),
7) submiting the original vimeo URL to universalsubtitles.org,
8) uploading the SRT to universalsubtitles.org.
I should add that whist the subtitles for this last video have been proof-read to produce high accuracy, useful searchable, and translatable subtitles; using the raw output of uweScribeBatchTranscription will produce results with higher accuracy than that of youtube due to the recognition engine using a personalised voice profile.
(see also: my post "workflow for educators" for more details about our speech recognition tools and workflow)