Hey everyone, hoping to get a sense of the real-world API usage here. I'm sketching out a cost projection for a new customer voice synthesis project using Cartesia, and the biggest variable is obviously the number of API calls we'll need to make.
I know the docs give you the raw technical costs per call, but I'm more interested in the practical workflow. For a single customer record in, say, a dynamic notification system (transaction alerts, personalized updates), how many calls are you typically making?
Here’s my initial thinking per customer:
* **One call for voice generation** (the core synthesis of the text).
* **Potential second call for voice cloning** if we're using a custom voice per customer (which seems powerful but adds a step).
* **Maybe a third for storage/caching?** If you're saving the audio output to your own CDN to avoid re-generating identical messages.
Does that align with your builds? Or are you finding ways to batch or cache more efficiently?
Also, has anyone done a comparison between using a unique voice per customer versus using a few "brand" voices and just personalizing the content? The cost difference could be significant at scale.
I’m coming from a heavy Looker/Tableau background where we optimize query costs, so I'm trying to apply that same mindset here. Any hands-on experience or gotchas would be super helpful!
Cheers, David
Data doesn't lie, but dashboards sometimes do.
Oh, I'm right there with you on this. We're planning a similar system for support call follow-ups, and the API call count was my biggest head-scratcher too.
You're spot on with the one-for-generation, one-for-cloning breakdown for a custom voice workflow. One thing we realized is that the voice cloning call isn't per notification, it's per customer, at least initially. You'd make that call once to create their voice model, and then just reference the `voice_id` in every generation call after that. So the call volume for cloning is way lower than for generation, which helps the math a bit.
Your third point about caching is the real key though, I think. We're definitely planning to store the final audio files. But we're also thinking about caching the actual text-to-speech result at the text level, if that makes sense. Like, if the phrase "Your order is confirmed" is said in the same voice to 100 customers, could we generate it once and reuse it? Not sure if that's possible with their API yet.
I'm also super curious about your last question on brand voices vs. unique voices. Has anyone run the numbers on whether customers even notice or care about a truly unique voice versus a really well-chosen generic one? The cost difference feels massive.