Experiments in AI
Regular readers of FITK (is there such an animal?) will have noted a
I used the program Suno to generate songs from poems and stories that have been published here over the last 20 years. This is an ongoing experiment. As my dealings with Chat GPT and other online AI services have been generally awful, I was pleasantly surprised by the quality of the music generated by Suno. Supply a lyric and a genre, press the ‘create’ button and ten seconds later you have two fully rendered songs. There were many failures of course—only about 5% of all attempts I would deem successful—although a couple did work the first time. I used Suno V3.5 for most of these; they have a newer version (V4) that is actually worse!
Before I go into the making of the songs, I would like to talk about AI: its definitions, its problems, its promises.
In its simplest form, a series of suggestions (prompts) is inputted to a program. I have experimented with AI image creation for a couple of years and found it to be useful, not necessarily to create great art, but to create illustrations that are simplified abstractions of concepts that actually work better to support a post or essay. The image adds a dimension to but remains subservient of the intent of the piece. An extremely arresting image is sometimes just too much and usurps the intent of the post. In one of the music videos I created AI-generated images to represent people. It would be difficult, and expensive to use real people as models. This use of AI is a real threat to commercial artists and photographers and has already devastated that industry. For my use I’m not making glossy magazine articles for mass media fare—this is a blog with about thirty readers MAX. Many of my photographic illustrations here have been AI augmented, not in the sense of creating false content, but for subtle enhancement of the photographic process. I use Topaz Photo AI™ to reduce noise, sharpen, and otherwise enhance the image quality, especially of highly-cropped images from my small-sensor camera (Pentax Q-7, before and after):
Language-based AI efforts (LLMs) create a similar quandary. Many of YouTube's videos are already written from AI content, usually bland, often over-simplified, and disingenuous. Suno can generate lyrics from your prompts, and this can work well with free-form styles. Proper lyrical content produces better results as the better the poetry that goes in, the better the song that comes out, sometimes the results are astonishing. ‘Massaging the meters’—making small changes to the lines—can improve the songs, but that is not-AI, it is what human songwriters have done for centuries. Another way to make effective lyrics is to input straight prose, the results are somewhat flakier, but the right match of material and genre can also be remarkable. This use of AI is, to my mind at least, new and exciting, holding great promise. It is not the malicious copying of existing artists and their songs.
This post marks the end of my experiments in AI music for a while. While I have been pleased with my results overall, there is a definite limit as to how much I can take, much less impose on my readers/listeners. If all of the songs I made with Suno were to be put on a traditional LP or CD I would have 5 full-length albums of this stuff! I’ve already done an album of love songs; there may be more in the future.
In a related development, my other website, Laxness in Translation has been linked to by ChatGPT—I’m now the source material of an AI program!
0 Comments:
Post a Comment