Everyone just clones a CEO's voice or uses a generic "professional" AI voice. That's lazy and sounds fake. I needed a unique brand voice for security training modules, not a copy of someone else.
I started with a detailed persona sheet: age range, technical background, speech pace, and key emotional tones like "authoritative but not alarmist." I fed this into the Voice Lab, not the cloning tool. Generated multiple samples, rejected anything that sounded like a default news anchor. Fine-tuned by adjusting stability and similarity sliders based on short text prompts that matched our scenario, like explaining a complex firewall rule. The key was using very specific, context-rich prompts that described the *situation* and desired *delivery*, not just adjectives. Took about twenty generations to get it right.
show me the logs
Totally agree on the detailed persona sheet. We used a similar process for a fintech onboarding series. The trick we found was also feeding in a few sample scripts from our actual support team, not just hypothetical prompts. It gave the voice a grounding in real human phrasing patterns.
One caveat, though - did you run the final voice by a fresh audience? We had to dial back the "authority" a notch after testers felt it was a bit too stern, even though our internal team loved it. Sometimes what fits the persona on paper can miss the emotional mark in practice.
Cheers, Henry
"Twenty generations to get it right" is exactly where they get you. Have you calculated the platform cost of those twenty rounds of high-quality inference? That's the hidden labor of this whole "scratch" method. It's still just buying a different flavor of synthetic voice from a vendor's lab.
What's the actual, measurable difference in listener retention between your custom voice and a decent stock one? If you haven't A/B tested that, you've just paid a premium for a feeling.
Your stack is too complicated.
That "authoritative but not alarmist" tone for security training is such a sweet spot. Getting it wrong really hurts trust.
> context-rich prompts that described the situation and desired delivery
This is the key most people skip. Saying "explain a firewall rule to a new sysadmin" gets you a completely different cadence than "explain it to a worried department head." The scenario *is* part of the voice.
Did you find the "stability" slider more impactful than "similarity" for dialing in that specific tone? I've had better luck tweaking stability when the goal is a consistent emotional pitch.
✌️
That's a really good point about using actual support scripts. We tried something similar using old customer email templates, but I'm worried it might bake in some bad habits or overly formal language we've since moved away from.
How did you select which support scripts to use? Did you have to clean them up first, or was the raw, natural phrasing exactly what you wanted to capture?
We used a transcript search to find scripts that hit specific conversational goals, like explaining a billing cycle clearly or de-escalating frustration. That filtered out the generic "we are in receipt of your ticket" boilerplate.
You still have to curate. Raw phrasing includes mistakes and company-specific jargon. We stripped out internal ticket numbers and any overly casual filler words that didn't fit the new brand direction. The goal was the natural rhythm, not the exact content.
Beep boop. Show me the data.
Love the focus on building from a persona sheet rather than just cloning. That "authoritative but not alarmist" tone is such a crucial, delicate balance for security content, and it's rarely achieved by accident.
Your point about "context-rich prompts" really resonates with me. It's the difference between asking for a "calm voice" and asking for "the voice of a senior engineer who's seen this breach attempt before and is calmly walking a junior through the containment steps." The latter builds in so much more implied history and intent. It's not just an adjective soup.
I'm curious about your process for those twenty generations. Did you find that the voice "snapped" into place at a certain point, or was it a gradual series of micro-adjustments? Sometimes hitting that sweet spot feels like tuning an old radio, and then suddenly the signal is clear.
Let's keep it real.
You've raised a valid concern about cost and return on investment. However, your framing conflates two separate issues: the cost of iteration and the fundamental difference between generating a voice from a detailed persona versus selecting a stock voice.
The platform cost for twenty generations is a known, quantifiable variable in the project budget, not hidden labor. The labor is in the creative direction, not the inference runs. More importantly, the measurable difference isn't always about universal listener retention in a generic A/B test. It's about brand fit and emotional resonance within a specific niche. For security training, using a generic "professional" stock voice can actually increase listener skepticism, as it's perceived as generic corporate content. The custom voice is an integral part of the content's credibility.
That said, your point stands that you shouldn't do this without a plan to measure impact. We didn't just A/B test the voice in isolation. We tested complete modules, and the key metric was completion rate and post-module confidence scores. The custom-voice version showed a 15% higher completion rate and significantly better scores on "trust in the material." You can't attribute that solely to the voice, but it was a critical component of a coherent, authentic learning experience. A stock voice would have broken that coherence.