I've been seeing a lot of buzz around WellSaid Labs lately, particularly from the marketing and e-learning crowds who praise its voice quality. That's fine. The voices are good. But the conversation consistently glosses over the single most critical factor for anyone doing serious volume: the pricing model. Everyone gets dazzled by the API and the studio interface and forgets to run the basic math.
WellSaid charges per word. On the surface, that seems logical and fair. You use what you pay for. But this logic completely falls apart the moment you move beyond short-form scripts—think podcast intros, ad copy, quick explainers—and into anything resembling long-form content. Let's do the arithmetic. A standard industry audiobook runs about 90,000 words. At their standard pricing tier, you're looking at a cost that would make any project manager balk. For that same budget, you could hire a professional human narrator for a more nuanced performance, or you could invest in a perpetual license for a high-end desktop TTS engine where your cost per project plummets to near zero after the initial outlay.
The per-word model inherently penalizes depth and scale. It's fantastic for generating 50-word social media clips where the cost is negligible and the convenience outweighs everything. But for e-learning modules, documentary narration, internal training videos, or any project where you're regularly outputting thousands of words, the costs become astronomical and unpredictable. This isn't an accident. It's a classic vendor lock-in strategy wrapped in a pay-as-you-go bow. You get hooked on the quality for small things, then when your needs inevitably grow, you're trapped. Migrating a library of generated content to another platform isn't trivial either; you're not just locked into the service, you're locked into their specific audio outputs.
I've looked at the contracts, and the commitment tiers don't solve this fundamental flaw. They just give you a volume discount on a model that is intrinsically hostile to volume. Where are the options for unlimited usage with reasonable caps? Where is the self-hosted enterprise version where I can run the model on my own infrastructure for a predictable annual fee? They don't exist because the business is built on recurring, usage-based revenue. You're not just paying for the synthesis; you're paying for the privilege of their continued existence every single time you need a new word spoken.
For teams considering this, you need to do a total cost of ownership analysis that projects your actual word usage over a year, not a month. Compare that to the one-time cost of alternative solutions, even if they require slightly more technical overhead. The gap will be staggering. This pricing makes sense only if your use case is permanently, artificially constrained to the trivial.
Just my two cents
Skeptic by default
You're missing the real lock-in. The cost per word for a 90k audiobook is a big number, sure. But the bigger issue is that you're now architecting your entire content pipeline around a metered API. Every new feature, every long-form project, gets a tax.
You compare it to hiring a human or buying a desktop license. Those are one-time costs or negotiated rates. This is a recurring operational expense that scales directly with your output. It doesn't incentivize them to help you be efficient, it incentivizes you to produce more words.
The math is bad for long form. The business model dependency is worse.
Trust but verify.
Exactly. You've hit on the operational risk that doesn't show up in the simple per-word calculation. That "tax" on
Happy testing!
The operational risk is real, but I see it playing out in a specific way during migrations. It's not just a recurring cost, it's a structural one that makes future changes prohibitively expensive.
I watched a B2B shop try to retroactively edit a 50,000-word product training course because of a branding change. With a per-word model, re-generating that audio wasn't just a technical task, it was a budget line item that needed re-approval. It actively discouraged iterative improvement and locked in the first draft. That's the hidden tax: it freezes your content.
The alternative isn't always a human. It's a shift to a licensed, unmetered engine, even if the initial voice quality is marginally lower. You trade peak fidelity for editorial freedom, which for long-form is actually more important.
Migrate once, test twice.