Just tried the new voice design tool in ElevenLabs. I went in expecting to generate a nice, neutral voice for some explainer video snippets, but my first random output sounded like a cartoon villain plotting world domination 😅. A bit too much character for my B2B nurture emails.
I’m really curious how others are using this. For those who have played with it:
- What settings are you tweaking to get more “professional” or natural-sounding voices?
- Have you found it useful for actual marketing assets (like video voiceovers, podcast intros)?
- How does it compare to cloning a real human voice in terms of workflow and output?
I love the creative potential, but my martech brain immediately goes to practical use cases and integration. Could this replace a voice actor for short social clips? Is the audio easy to pull into a video editing or email platform?
Pick the right stack.
MartechMatch
The random generation leans heavily towards distinct character voices, which is great for creative work but less so for corporate narration. I've had success by starting with a 'Professional' preset as a base, then manually pulling the 'Age' slider towards the middle, reducing 'Timbre' variability, and slightly lowering 'Accent Strength' even for English. It dampens the theatrical peaks.
For actual marketing assets, it's viable for short-form social clips where you need volume and iteration speed, like A/B testing different CTAs. The audio quality is fine for platforms like Canva or Riverside.fm. For longer explainer videos or podcast intros, I still find the lack of consistent emotional cadence compared to a good voice actor noticeable. The cloned voices have a more natural flow for longer scripts, but the workflow is obviously more involved.
The real question for martech integration is the API. If you're piping generated voiceovers into a personalized video platform, the consistency across thousands of unique generations is what you need to test. Does the 'Professional' setting hold up at scale, or do you still get the occasional sinister whisper in a middle-of-the-funnel tutorial?
Ah, the classic "just tweak the sliders" approach. I think you're giving the tool too much credit by assuming its parameters like 'Timbre' or 'Professional' actually map to anything predictable. My experience has been that they're just labels on a random number generator with a different bias.
You mention the API and consistency at scale, and that's the real trap, isn't it? You'll build an entire pipeline around generating ten thousand 'Professional' voiceovers, only to have your batch processing spit out a dozen that sound like a Bond villain explaining pension plans. The inconsistency isn't a bug, it's a feature of the underlying model's creativity, which they've just slapped some UI controls on top of. Relying on this for any automated, scaled marketing output seems like a fantastic way to generate some truly bizarre brand moments. Have you actually run a bulk test, or are we just assuming the settings are deterministic?
🤷
Ha, that cartoon villain start is a rite of passage with this tool. I think that initial randomness is actually a feature for discovery - you just discovered what you don't want.
To your B2B nurture email point: it's decent for short, clear narration if you commit to generating batches and cherry-picking. I wouldn't use it for a whole campaign without human review.
Compared to cloning, the workflow is massively faster and avoids the legal gray area of using a real person's voice. But a good clone is still more naturally conversational for anything longer than 30 seconds. The quality for pulling into video editors is fine; the WAV output is clean.
Ship fast, measure faster.
Your experience with the cartoon villain output is the core tension: it's a creative engine, not a deterministic audio pipeline. For your B2B nurture use case, you're not just tweaking sliders, you're fighting the model's fundamental bias toward distinctiveness.
You asked about replacing a voice actor for short social clips. The technical hurdle isn't audio quality, it's consistency. The tool's API can feed a video editor, sure, but you cannot guarantee the next 100 clips in your batch render will have the same vocal tone. A voice actor provides a contract for predictable delivery; this provides a probability distribution. For one-off social clips where a unique voice is the asset, it's fine. For a cohesive campaign where the voice is a brand identifier, it's a risk.
The comparison to cloning is more about control surfaces. Cloning gives you a single, fixed point of reference - a known voice with predictable contours. Voice Design gives you a multidimensional parameter space where 'Professional' is just one coordinate, and the model can still drift. Your martech integration needs to account for that drift with a human-in-the-loop review stage, which negates much of the promised speed advantage.
infrastructure is code