我打造 Resona 按用量付費語音克隆平台的心得

What I Learned Building a Pay-Per-Use Voice Cloning Platform

Every AI voice tool I tried had the same shape: pick a plan, $10–30/month, cancel anytime (you won't). Before you even know if the voice quality works for your project, you've committed to a month. I kept thinking: this pricing model fits SaaS dashboards, not creative tools you use in bursts.

So I built Resona the other way around — credits instead of a plan.

The model, in one paragraph

You buy credits ($1 = 50 credits) and they never expire. Cloning a voice costs 100 credits. Generating speech costs 1,500 credits per 1M UTF-8 bytes of text — which works out to fractions of a cent for a typical paragraph. If you clone one voice and narrate one video this month, you pay for exactly that. If you don't touch it for two months, nothing is wasted.

Three things I learned shipping it

1. Sample quality beats sample quantity. People email us asking why their clone sounds off, and it's almost never the algorithm — it's a phone recording with room echo. Thirty seconds of clean audio from one consistent mic beats five messy minutes. Record in a soft-furnished room, keep the mic distance fixed, and don't mix takes from different days.

2. Billing by bytes changes how people write. When you pay per minute of studio time or per month of access, nobody thinks about the text. When you pay per byte, people notice that "um, so, yeah, basically..." costs money. The side effect is better scripts — tighter sentences, fewer filler words. I didn't expect a pricing model to act as an editor, but here we are.

3. Multilingual is not a checkbox. Resona generates speech in 83 languages, and the requests we didn't anticipate are the fun ones: the same cloned voice reading a script in English and then in Japanese for a localized tutorial. Keeping a voice "recognizably the same person" across languages turned out to be the feature people care about most — more than raw audio fidelity.

What's next

Voice design (describing a voice in text and generating it), transcription with speaker separation, and a public library of 200,000+ ready-to-use voices. The roadmap is driven by what burst-users ask for, because they're the ones the pricing model was built for.

Full disclosure: I'm the founder of Resona (resonaio.com), so obviously biased toward pay-per-use — but the per-byte vs per-month math above holds no matter which vendor you pick. If you try it, I'd genuinely like to hear where the model breaks for your use case.