dictating new wordage with Dragon
Aug. 13th, 2012 10:11 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
julesjonesFriday night a week and a bit ago I felt up to experimenting with using Dragon to write new fiction material. I specify fiction, because a lot of my blog posts over the last few months have been written at least partly with Dragon, and an awful lot of my typing at work has been with Dragon rather than keyboard. But fiction's different, for various reasons.
The current primary WIP is one that I wouldn't even attempt to use Dragon on. There's too much staring into space while I turn shape of story into words, and dealing with the error correction would thoroughly disrupt my train of thought. But occasionally I have a chunk of text right there in my head, and that might be more amenable to scribbling it down quickly via dictation rather than via keyboard or handwriting recognition (I still carry around a Palm IIIxe for the latter purpose, even if it doesn't get much use nowadays).
That still wouldn't work for anything longer than one or two sentences using my normal mode of dictating, which is to dictate a sentence or two and then do the corrections. But Dragon can also transcribe from a sound file with previously recorded dictation, and my shiny new headset came with sound recording software.
I had something that wasn't full on "the text is in my head and I simply need to type it before I forget the words", but did have a reasonable chance of me being able to witter on for more than one sentence before having to stop to formulate the next bit. That makes it feasible to dictate a passage, and then go through the complete transcription file afterwards to clean up the errors by listening to the recording and making corrections. (Cleaning up the errors via Dragon is essential to improve recognition in the future.)
Alas, I'd left it a bit too long before doing something with it, and some of the words had escaped. There was much pausing to think of how the next bit should be phrased, and thus many points where I paused the recording in order to stare into space.
The actual dictating into recorder bit -- not entirely a success. It made me much more consciously aware of how much I sit and spare into space, for one thing. :-) But the two big problems I consciously noticed were that I find it much more difficult to compose the next sentence without the visual feedback of the previous sentence on the screen/page in front of me, and that I still find it disruptive to my train of thought to have to explicitly dictate punctuation. However, the punctuation problem is, I think, less noticeable when doing this sort of "dictate for later transcription" than when doing real-time dictating/correcting.
The sound file I ended up with is 9m38s long, according to Audio Commander. That's with a fair bit of hitting the pause button when I needed to stare into space, and I'm not quite sure how long I was actually dictating for, although I doubt it was longer than 20m.
The raw transcript has a word count of 829 words. Many of those words are not what I actually said. Some of them are words that I did in fact say, but intended to be dictation commands rather than text. The transcription is actually pretty appalling, compare to how I'm getting on with using Dragon at work to dictate letters etc. :-) This may be partly because I didn't remember to RTFM before starting this lark, and thus didn't set up the digital recorder as a separate source. I did shift the accuracy-speed balance all the way over to accuracy, so that's not the problem. Or it may be that I haven't written any fiction with this profile, so its sampling of my likely pattern of word usage is off.
It's taken about 20 minutes to go through the transcript playing back the audio and correcting it where necessary. That's just the first pass, to correct the transcript to what I actually said, glitches and all. This is necessary to teach Dragon to adapt to my speech patterns -- if I don't correct it accurately to exactly what I said, rather than what I meant, that will decrease its recognition accuracy in future.
Next pass will be to go through and take out the commands it mis-recognised as text, and do the corrections I was trying to do with those commands (typically deleting a few words because I'd changed my mind about what I wanted to type). And after that I'll do a revision pass.
So, slow and tedious. But doable, and could be useful in certain circumstances. I might put the raw transcript and the various editing passes online, if people are interested in seeing what sort of errors Dragon is making.
The current primary WIP is one that I wouldn't even attempt to use Dragon on. There's too much staring into space while I turn shape of story into words, and dealing with the error correction would thoroughly disrupt my train of thought. But occasionally I have a chunk of text right there in my head, and that might be more amenable to scribbling it down quickly via dictation rather than via keyboard or handwriting recognition (I still carry around a Palm IIIxe for the latter purpose, even if it doesn't get much use nowadays).
That still wouldn't work for anything longer than one or two sentences using my normal mode of dictating, which is to dictate a sentence or two and then do the corrections. But Dragon can also transcribe from a sound file with previously recorded dictation, and my shiny new headset came with sound recording software.
I had something that wasn't full on "the text is in my head and I simply need to type it before I forget the words", but did have a reasonable chance of me being able to witter on for more than one sentence before having to stop to formulate the next bit. That makes it feasible to dictate a passage, and then go through the complete transcription file afterwards to clean up the errors by listening to the recording and making corrections. (Cleaning up the errors via Dragon is essential to improve recognition in the future.)
Alas, I'd left it a bit too long before doing something with it, and some of the words had escaped. There was much pausing to think of how the next bit should be phrased, and thus many points where I paused the recording in order to stare into space.
The actual dictating into recorder bit -- not entirely a success. It made me much more consciously aware of how much I sit and spare into space, for one thing. :-) But the two big problems I consciously noticed were that I find it much more difficult to compose the next sentence without the visual feedback of the previous sentence on the screen/page in front of me, and that I still find it disruptive to my train of thought to have to explicitly dictate punctuation. However, the punctuation problem is, I think, less noticeable when doing this sort of "dictate for later transcription" than when doing real-time dictating/correcting.
The sound file I ended up with is 9m38s long, according to Audio Commander. That's with a fair bit of hitting the pause button when I needed to stare into space, and I'm not quite sure how long I was actually dictating for, although I doubt it was longer than 20m.
The raw transcript has a word count of 829 words. Many of those words are not what I actually said. Some of them are words that I did in fact say, but intended to be dictation commands rather than text. The transcription is actually pretty appalling, compare to how I'm getting on with using Dragon at work to dictate letters etc. :-) This may be partly because I didn't remember to RTFM before starting this lark, and thus didn't set up the digital recorder as a separate source. I did shift the accuracy-speed balance all the way over to accuracy, so that's not the problem. Or it may be that I haven't written any fiction with this profile, so its sampling of my likely pattern of word usage is off.
It's taken about 20 minutes to go through the transcript playing back the audio and correcting it where necessary. That's just the first pass, to correct the transcript to what I actually said, glitches and all. This is necessary to teach Dragon to adapt to my speech patterns -- if I don't correct it accurately to exactly what I said, rather than what I meant, that will decrease its recognition accuracy in future.
Next pass will be to go through and take out the commands it mis-recognised as text, and do the corrections I was trying to do with those commands (typically deleting a few words because I'd changed my mind about what I wanted to type). And after that I'll do a revision pass.
So, slow and tedious. But doable, and could be useful in certain circumstances. I might put the raw transcript and the various editing passes online, if people are interested in seeing what sort of errors Dragon is making.
(no subject)
Date: 2012-08-14 05:34 am (UTC)Oh yes, please!