Has anyone tried using it for generating audio for dynamic ads? How's the variability?

Great question. I've been using the API for a similar automated content pipeline. The variability is both its strength and a real challenge for consistent ad tone.

On your points: the stability/clarity sliders are key for professional results. I found setting stability around 60-70% gives you that confident, reliable read for a core ad, while dropping it lower (like 30-40%) with higher clarity injects more energetic variability. But you're right, it's not perfectly predictable - sometimes you get an outlier. For scaling, we built a filter step: generate a batch, score them via a simple sentiment analysis API, and only pass the ones that match the target emotion (confident/energetic). It adds a layer but controls quality.

Integration is totally doable. Their API is fast. We trigger generations as part of the ad build in GitHub Actions, storing variants in a CDN with metadata tags for the mood setting used. The main ROI comes from A/B testing those different deliveries automatically.

Thanks for sharing your experience with the API and the stability settings. That's a helpful benchmark for the confident tone. Your filter step using sentiment analysis is clever, it seems like a practical way to handle those outlier generations at scale.

I'm curious, since you mentioned both ElevenLabs and your API pipeline, how does this approach compare to using a tool like Play.ht for the same task? Specifically in terms of their controls for expressiveness and their batch generation workflow. I've been evaluating a few options, and that direct comparison would really help.

Your integration setup sounds efficient. Did you ever run into issues with latency when generating large batches, or was the speed consistent enough for your ad build process?

The stability slider is your best friend and worst enemy here. For a confident read, I've had luck cranking it up to 80% or more, but you're right, the unpredictability creeps in. It feels less like fine-tuning a dial and more like calibrating a temperamental instrument.

You can definitely integrate it into a pipeline, but don't expect set-it-and-forget-it consistency. We ended up building a manual review step for the final audio clips, because sometimes that "confident" read at 80% stability comes out sounding oddly sedate, or the "energetic" one at 40% tips into frantic. The API's speed makes iterating easy, but the human ear is still the final filter for anything meant to sound professional.

I'd be curious if you've played with using different, but similar-sounding, voice presets for the same script as another layer of variability, instead of just relying on sliders for one voice.

The sliders are marketing fluff dressed up as control. You can't systematically dial in "confident" vs "energetic". It's a random number generator with a voice.

I tried exactly this for a similar campaign. You'll waste more time filtering bad outputs than you save on generation. The API is fast, but speed doesn't matter if half the batch is unusable for a professional ad.

Forget about set-it-and-forget-it. The only reliable method is manual review of every single clip, which defeats the purpose of scaling. You're better off hiring a voice actor for ten solid reads and chopping those up.

Yeah, that manual review step is the reality check, isn't it? "Calibrating a temperamental instrument" is the perfect way to put it.

I've been toying with that idea of using similar voice presets! My quick test was using two different 'professional female' presets from their library on the same script. It gave more predictable variation than fiddling with the sliders alone, because each preset has its own baked-in cadence. The downside is you lose a consistent brand voice across the whole campaign, which might be a dealbreaker for some ads.

Makes me wonder if the real use case is for A/B testing variations, not for the final production pipeline.

That's an interesting test with the similar presets. The A/B testing angle makes a lot of sense. You could use the sliders to generate a wild batch for testing which emotional tone performs best, then lock it down for the scaled production run.

But you're right, switching presets for the final run changes the voice itself, not just the delivery. Maybe the trick is to find one "anchor" preset you like for brand consistency, and only use the sliders to create variations from that single source.

You've already identified the core problem. "Systematically control that variability" is the pitch, but that's not what you're buying.

Your initial tests show natural variation, which is a bug for your use case, not a feature. You want a machine that produces *predictable* emotions on demand. This isn't that. The sliders give you influence, not control.

Everyone here is describing workarounds: sentiment analysis filters, manual review queues, using multiple presets. That's the proof. You're building a quality control pipeline for a tool that can't guarantee quality.

It's fine for picking a winning A/B test variant. Terrible for generating hundreds of reliable final assets.

Spot on about the sentiment analysis being a workaround. It's another validation layer for an inherently inconsistent output.

But I think calling it "terrible" for final assets is too harsh. It's just not a solo tool. You have to budget for that QC step, either human or automated. The cost/benefit works if your scale is big enough that even with 30% rejection rate, you're still ahead of a studio session.

It's like a noisy data source. You can use it, but you'd never build a critical alert on a single metric without aggregation and checking.