Not all experiments work
I had plans. With the next season of Grim Khonsu I’d add audio. I’d heard of Spoken, a website/company that allowed users to produce multi-voice AI audiobooks, and this sounded ideal. I’d prefer to use human narration, but I’m working from a limited budget, and I don’t have the voice to do it myself, so I’ll explore AI options.
My plan involved releasing each episode of Grim Khonsu as text and audio, with the audio files going on YouTube a few days later.
But would the plan work? Was it viable?
I needed to test it.
For a while now, I’ve been toying with the idea of interspersing the longer Grim Khonsu tales (the ‘seasons’) with shorter stories that run over two or three weeks. I had an idea for such a story, so I used this to test my audio plan.
I wrote the story and uploaded it to Spoken. They provide far more customisation of voices than other tools I’ve explored. It’s possible to alter speed and add ‘emotions’ to different passages.
The voices are better than others I’ve heard, too. They’re a mixture of purely synthetic and human clones, and Spoken pulls these voices from a few sources, including ElevenLabs. With human voice clones, it’s all done through agreements, so they’re not ripping any of their human talent off (at least, as far as I’m aware).
One of my first decisions was around accent. As a Brit, my first thought was to use British accents. But detective noir is strongly associated with America. My dilemma — go with what I wanted, or let the audio meet the most common expectations?
To answer this, I returned to my reasons for writing Grim Khonsu. A big part of it is the enjoyment of the work — writing, getting to know the characters and story, moulding the early draft into a stronger product. Another part of working on Grim Khonsu is to better understand my larger story-universe (I’ve commented before how this serial is one strand in a larger space-opera story). But a third factor with Grim Khonsu is that it’s free. It’s a low-cost (nothing financial, but there’s a temporal cost in reading a story) introduction to my writing, and I’m hoping that some readers will be enticed into exploring more of my books.
This final aspect was important in my decision. If Grim Khonsu is a calling card, it has to appeal to the largest demographic of potential readers, and they’d be expecting American accents.
With voices for each character selected, I worked through the text, listening to the audio, making adjustments where needed. I ended up with two short episodes (about eight minutes each). Perfect.
Only it wasn’t.
I waited a couple of days, then listened to the downloaded files. And I wasn’t happy with what I heard.
Oh, the quality of the voices was great (for AI). There was some articulation present, and while those ‘emotion’ prompts hadn’t made huge differences, they did lift certain lines.
But it still felt stilted. I aim for a certain type of style when writing Grim Khonsu, influenced by Raymond Chandler. Noir leans into ‘voice’, and this AI narration didn’t have it. On the page dialogue is a sparring match, but the audio version sounded like two people in separate booths, each reading their script.
In short, it lacked the life I felt it needed.
I was disappointed. Of course I was. Producing an AI narration isn’t a case of clicking a button and it’s done. I’d spent time on this. I didn’t like to think of that time being wasted.
I had options, though. I could’ve returned to Spoken, reworking the audio — different voices, going deeper with the speed and emotion controls. But I don’t have unlimited time. I’m already behind where I wanted to be this year. I couldn’t justify spending more time on this audio, especially when there was no guarantee the results would be significantly better.
I had those downloaded files, of course. I could’ve gone ahead with them, maybe explaining that they weren’t as good as I’d anticipated. But I don’t like putting out material I’m not happy with. I aim to make everything I write (and by extension narrate) as good as it can be. Nothing will ever be perfect, but that doesn’t mean I can let low quality slip through.
So the plan has failed. I won’t be releasing an audio version of Grim Khonsu at the moment. In the future, maybe — if AI tools reach a standard I’m happy with for this particular serial, or if I can afford decent human narration. But for now, Grim Khonsu is text-only.
I’m still in the early stages of the second season (although I have a firm grasp of the story and characters, so I expect to make good progress soon). But I do have that short story.
This means that Grim Khonsu is returning. I’ll publish the first part of this interlude next Thursday, with the concluding part the week after. As usual, it’ll be free to read. But no audio.
Except the one that plays in your mind as you read. I don’t think AI will ever improve on good old human imagination.