Further Adventures in AI
I've been experimenting with AI over the past few months. No, I haven't been setting up my own Artificial Intelligence powered by multiple server farms and using the collected content of The Sleaze as a database. Rather, I've been playing with various so-called AI tools available online in order to see if I could find any application for them or, indeed, if they were actually any good. The earliest fruits of this dabbling were a couple of podcasts I produced where all of the voices were provided by AI based TTS services. I've referred to these before and to recap, the main tool used was Google's Notebook LM, to which you can input sources in a variety of formats and it will produce an analysis of them, along with an AI generated conversation between two synthesised characters, going over the main points and arguments arising from the sources. The voices and their conversation are uncannily realistic and their discussions of various old stories from The Sleaze, treating them as if they were serious news reports, is quite fascinating and invariably, unintentionally, hilarious. The other voice tools I used required me to provide scripts for them to read, with the results being somewhat variable. Straightforward TTS AI voices are undoubtedly far more realistic sounding nowadays, with more natural sounding intonations but nonetheless still have a tendency to sound somewhat stilted in their deliveries.
Another TTS service I used allows you to use various celebrity voices, which can vary enormously in quality, to read your scripts. As the free version only allows you to stynthesise a few seconds of speech at a time, a lot of editing together of these fragments is required. The biggest problem I've found with these is that they invariably seem to speak too quickly and I've found myself having to drastically reduce their tempo during the editing process. The results, though, can be quite effective. I've mainly used them to produce brief inserts - usually featuring Donald Trump or Jerry Springer - which separate the main segments of my regular podcasts. While all of the AI powered audio tools I've used so far have proven useful and have brought something new to my podcasting, they're still pretty limited in what they can do, still requiring an enormous amount of input from the user, so they can hardly be said to be time saving with regard to the creative process. In addition to these, I've also been using AI to generate the images that accompany new stories on The Sleaze. While these do save me time and provide unique images in some way reflecting the stories' themes, actually getting something relevant can be an uphill struggle, requiring multiple attempts using variations on the original cue. Some of the results seem pretty wide of the mark. Take, for example, this one, which was one of the images generated for the recent 'Rise of the Clockwork Nazis' story:
A striking image, to be sure, but not at all relevant to the simple prompt of 'clockwork Nazi'. Now, the prompt that the AI service attaches to the picture is this:
"A damaged clockwork Nazi dog patrols a deserted factory at night, its gears grinding as it searches for intruders in the shadows of a post-apocalyptic industrial wasteland."
Which is a long way from the actual two word prompt, which makes no mention of dogs, or post-apocalyptic landscapes (plus, to be a pedant, there is no indication in the picture that the dog is a Nazi). So where does the additional information come from? Is this another case of AI making stuff up?
Anyway, while currently available AI tools are all very interesting and have their uses, it is still very limited, both in what it produces and how it interacts with humans, seemingly still having real problems in actually comprehending what is being asked of it. But hey, I'll still go on using it, even though I feel somewhat guilty at the amount of energy it is probably using just to perform these frivolous tasks for me. Then again, I don't fly, so I guess that more than offsets the environmental damage.
0 Comments:
Post a Comment
Subscribe to Post Comments [Atom]
<< Home