Hi,
I'm not sure if this thread is really about the accuracy of the speech-to-text process, but because Frank_xrx asked, I thought I might give my own answer. I've been working with this workflow for the last four weeks now, and have finally decided that it is not possible to do serious work with the tools as they are at the moment.
All the comments from Adobe about the quality of the voice-recording affecting the accuracy of the text analysis are true, but, as others have noted, when your sound-track has spontaneous speech you cannot always separate one voice from the background, nor have recording-studio quality all the time. In such cases the text that is produced will simply not be accurate enough to use, whether you have attached an accurate script or not. Also, this speech analysis process does not like long pauses between speakers (e.g. if you have any action in your video, this will really mess up the timing of text that is produced).
However, above and beyond all this, the main flaw in the whole process, from my point of view, is that even with a word-perfect script typed into Story, attached to video clips in On Location, and then "speech-analyzed" in Premiere... if the resulting text has omissions or mistakes, there is almost nothing you can do to edit them. OK, I know there is an editing facility but it is a sham. If you add missing words into the text, then these words appear in the text, but they never get a time stamp, so they never appear in the exported xml-file, and they can never be used to "go to" a particular point in the video. If, like me, you want to find out where and how many times certain words are used in your video, then none of these newly added words will be found! Also, if you look closely, you'll find that you cannot add anything before the first word of the analyzed text that is produced in Premiere, nor can you add anything after the last word of that text. It is just not possible (right click on the first and last words to see the greyed out items). I've had several instances where the first speaker in a scene is completely missed out of the analyzed text, and also, in my experience, the analysis process always (i.e. always) misses off the last word in each scene, regardless of sound-track quality or the attachment of a full script. When these things happen this is a major problem for me. For me, the feature of searching for and playing back individual words or phrases in a video was the main attraction of the Production Premium Suite; this was why I bought it. I'm trying to produce spoken and visual "concordances", as a step up from the text-only versions now available (sorry if that's a bit esoteric).
I know that a lot of editing can be done to the analyzed text if you use After Effects. And this is great if you want to show subtitles. But anything you add, or change in AE is still beyond the reach of the xml files exported by Premiere or Encore, so, again, no changes you make in AE will be searchable (or findable) in the final video. There is another route to producing a text-searchable video, which involves putting physical cue points into the video whilst in AE, and then writing a Flash program to use those cue points, instead of using the xml-file data. I haven't tried that yet, as I kept on hoping that the (frequently) advertised speech-to-text workflow would actually work. But after four, increasingly frustrating weeks I've given up on it, simply because I can't get even a small proof-of-concept project to work exactly as advertised. Starting a serious project, with the current tools, would be unwise.
I realize that there must be work-arounds for some of the problems, but at the price of this suite (even more expensive for those of us who don't live in the US), I feel that things should work as advertised. This is not Version 1.0 from a start up company.
So, I hope this wasn't too verbose. I have specific requirements of the speech-to-text workflow which I've tried to explain above. If your requirements are not the same, then maybe you can get away with the work-arounds.. and... this thread was originally about subtitles, and there is some hope for those, using AE. I thought I'd still write this, though, as I need a bit of an outlet for all these weeks of frustration :o)
|
