Arachne Audio and Localizing Covers (scaling day 3)

Cross-posted from AI Marketing for Storytellers.

So, trying to get perfect international audio generations has been just as thrilling as I expected it to be — and I’m not being sarcastic, we’re all working v v v v hard over here.

Let me talk you through it —

The problem with doing big chunks of audio in languages you don’t understand, is that you won’t catch prosody errors or drift — but with audio we have the benefit of working in English. You can’t really ask a model to translate from English to English to catch mistakes, but you can with mp3s!

We tried a version where we ran every sentence through and assembled them on the far end, but that wound up sounding very Frankenstien’s monster — each sentence had its own individual emotion and tone.

So now we’re hacking everything into paragraph levels, which keeps the emotional theme through line better, which is great, except for when a piece of your audio paragraph sucks.

Here’s the rubic we’re grading the audio against:

So now we’re working on a system to record first at the paragraph level, then do soundchecks on each paragraph essentially, and if a paragraph fails, to break it out at the sentence level, so we can re-record whatever needs re-recording, and then stitch it all back together again.

Luckily, I have mad experience with keeping track of data, from my translation vs of Arachne, so this isn’t really that traumatic, although I think you might enjoy seeing this mess of a file path we’re up to now via experimentation, lol:

Para-surgery indeed!

Currently we’re using 11labs API for the recording, and then Gemini Flash (I think? it’s a cheapy vs of Gemini) to listen to the mp3s and aribtrate, and then cycle through recording again, if needbe.

Is this going to work?

No clue yet, but at least my demo’s gonna be in English, so I can hear if it doesn’t.

I do think we’re making progress though — the goal is to make it SO SOLID in English that any other language is a slamdunk.

On all of the plus sides in the world though, I solidly finished my book last night, and then redid its cover with Jack before I fell asleep — and now we’re working on localizing the covers for all 50 of my translations:

We’re using the Noto family of fonts for this, because it can do all the CJK langs (Chinese, Japanese, etc) and Cyrillic ones, in addition to all of the ones in between (Arabic, Devanagari) and being free (thanks Google!)

I really like this final version of the cover, I think Jack knocked it outta the park — especially given that I needed the space for the alternate languages to live in.

While my audio tests churn, we’re also finishing localizing the language packs themselves for nonfic. I’ve got an instance of Claude Code hard at work with a fleet of worker bees, hitting all my lang folders, and then I’ll run the title and the blurb through Arachne to get the final version, so that I can run (maybe! fingers crossed!) my Kobo version tonight — or at the least assign all of these version an ISBN from Bowkers. (Did you know that Bowkers doesn’t take non latinate languages? THE BASTARDS, lol, but it is what it is, for those languages, the title of the book will just be “How to Think With AI — in Hindi” — the ISBNs are more for me than for anything else really.

Some other interesting considerations include profanity levels and — in some of these langs — the fact that I’m a woman writing rather quite boldly.

I don’t think I’m going to get in trouble personally — while I’d love the chance to visit all of the places I’m translating this book into, I don’t think I’ll get to live that long, and my chances of being extradited are low — but it’s worth One Serious Last Consideration once I’m out of the coding weeds tomorrow.

I already had to think about that a lot when I ran my Dark Ink Tattoo series through Arachne’s original 28 langs (including Arabic and both Simplified and Traditional Chinese).

Those books are, to put it lightly, HELLA GAY. (Like me!)

I decided that I’m not going to defang my fiction — and, while I might not publish them on Amazon.uae, because I do value my Amazon account, and I don’t want to get flagged — I am going to publish them everywhere else.

They probably won’t make me a ton of money, but there could be people from other langs who’re learning those langs who want to read — or diaspora readers — or actual queer people in those countries who want to read queer books about vampires on the sly.

It’s not for me to decide, it’s just for me to send my books out with an open heart.

I sat down last night and created a worker for funsies that can access Bookbub’s back end and apply for Bookbub Featured Deals for me, which is pretty great, because, again, Claude Code lets you time things now, so I can just have that on an auto-timer, and not think about it ever again in my life, huzzah! (Also that worker will never experience the, “I have done this ten times with no result, ugh,” shame that’s stopped me from doing it manually — in addition to my general lack of free time.)

Because, uh, I’ve been dodging all sorts of Real Responsibilities in regards to my writing career and ads maintenence, but more on that tomorrow, this post is already long enough!

xo!

Cassie

PS: some final thoughts from Codex, re: coding in general