ElevenLabs vs Synthetic Voices for documentary narration - which is more convincing?

Hello everyone, I've been lurking for a while but this is my first real post. I'm currently in the procurement phase for a significant documentary series project (budget is a major consideration, so TCO is critical), and the narration component has become a surprisingly complex vendor-benchmarking exercise. We're evaluating AI voice solutions for portions that require narration in different tones and languages, where hiring a human VO for every variant isn't feasible.

Our shortlist has essentially come down to ElevenLabs and a category I'll call "Synthetic Voices" (by which I mean more traditional, older-generation TTS services like Amazon Polly, Google's WaveNet voices, Microsoft Azure Neural TTS, and even some older offerings from IBM Watson). The core question for our team, which I'd love this community's insight on, is: **which produces a more *convincing* narration for a documentary context, where the voice needs to carry authority and subtle emotional weight, not just be clear?**

Here’s my hesitant, very detailed breakdown of our evaluation process so far. We've been testing with identical scripts (a 300-word segment on a historical event):

**ElevenLabs Pros & Cons (based on our tests):**
* The "realism" in terms of breathiness and natural pacing is, at times, stunning. The voices don't sound like they're on a metronome.
* The voice cloning feature is a potential game-changer for consistency if we ever wanted to mimic a specific narrator's style for spin-offs.
* However, we've noticed a critical pitfall: sometimes the inflection goes *too* dramatic or oddly placed on a seemingly unimportant word, breaking the "trustworthy narrator" illusion. It can feel like a performance, not a recounting.
* Pricing model is a big concern. The per-character cost for the highest quality tiers, especially for long-form content, could spiral compared to a more predictable synthetic voice subscription.

**Traditional Synthetic Voices (Polly, Azure, etc.) Pros & Cons:**
* The consistency is rock-solid. Once you find a voice and tune the SSML settings (for speed, pitch), it's perfectly reproducible.
* The pricing is often more predictable for high-volume usage, which appeals to our procurement side.
* The downside is the "synthetic ceiling." Even the best neural voices can have a tell-tale flatness or a slight "digital sheen" in sustained narration. They sound good, but you'd never mistake them for a human in a quiet, reflective segment of a documentary.

Our current dilemma is whether the occasional burst of breathtaking realism from ElevenLabs outweighs its occasional odd inflection and higher cost volatility, versus the reliable, slightly less convincing output from the traditional synthetic providers. Has anyone here done a direct, long-form A/B test for a serious project like this? Are we missing a key parameter in our evaluation? I'm particularly worried about audience immersion being broken by an odd vocal choice from an AI, which feels like a risk with both options, just in different ways.