TL;DR: David Pinto, our Senior Project Manager and AI expert, shares his experience building an AI version of himself. While creating a visual avatar was easy, developing a realistic voice clone took months of testing. He discovered that achieving an authentic voice requires diverse and natural training data. Using a professional voice clone ultimately made a significant difference in quality.
Jump to section:
Don’t have time to read the article? Listen instead:
Ready to explore AI avatars and voice clones for your business? Ask Iguazu
Why I tried to create an AI version of myself
AI avatars and voice cloning technology are rapidly changing how individuals and businesses create digital content. From training materials and personalised communications to scalable video production, AI-powered avatars are becoming a powerful tool for modern organisations.
But how easy is it to actually create a convincing AI version of yourself?
Over the past year, I set out to build an AI clone of myself using an AI avatar and a cloned voice. What initially seemed like a straightforward experiment turned into a long journey of trial, error, and experimentation.
Along the way I learned a lot about how AI avatars work, why voice cloning can be difficult, and what it really takes to produce a digital voice that sounds authentic.
How we made our AI avatars
The project began with creating the visual component of our AI clones. Last year, the Director of Iguazu, our Video Designer, and I recorded video footage to train AI avatars using an AI tool called HeyGen. The process involved filming ourselves delivering a set script so the platform could analyse our expressions, movements, and speaking patterns.
Once uploaded, the AI avatar tool trained a digital avatar capable of presenting video content that looks and behaves like the real person. The avatar creation process worked well and produced realistic digital versions of us. However, an avatar without the right voice still feels artificial.
What we learnt: To complete the AI clone, we needed convincing voice replicas.
First attempt: using instant voice cloning
To generate our AI voices, we used ElevenLabs and its Instant Voice Clone feature.Â
Our first attempt used the audio captured during the avatar recording session. The results were interesting.
Our Director’s voice clone was almost perfect on the first attempt. My other colleague’s voice was close but not entirely accurate. My voice clone, however, sounded nothing like me.
This highlighted an important reality of AI voice cloning: results can vary dramatically depending on the voice data and recording quality.
What my first voice sounded like. This didn’t sound like me at all.
Recording new audio for voice training
My colleague and I decided to record new audio samples specifically for voice cloning. Using the Instant Voice Clone again, we created fresh recordings designed to give the model better training data. This second attempt significantly improved my colleague’s voice clone, making it almost indistinguishable from the real thing.
Unlike my colleague, it was obvious that my voice still wasn’t convincing.
Multiple attempts to improve the voice clone
Determined to improve the results, I recorded my voice eleven additional times using the Instant Voice Clone process. Each attempt experimented with slightly different recording styles:
- Speaking more slowly
- Changing tone and pacing
- Recording in different environments
- Using varied topics and phrasing
While these experiments helped slightly, the voice still only sounded around 60–80% similar to my real voice.
Often it would sound somewhat correct, but at other times it would drift far away from how I actually speak.
My attempts at improving my first voice clone. It definitely improved, but still nothing like me.
Experimenting with voice metadata
Next, I explored whether metadata labels could help improve the AI model. I added the following voice attributes:
- Language: English
- Accent: Received Pronunciation
- Gender: Male
What I learnt: This provided some improvement, but the voice still lacked consistency and authenticity.
Testing accent data
For another experiment, I created a new Instant Voice Clone recording where I tried to sound as natural as possible and added different accent metadata:
- Language: English
- Accent: Portuguese
- Gender: Male
I chose Portuguese because it reflects my background and I was curious whether including that information might help the AI interpret my voice differently. Unfortunately, the improvement was minimal.
At this point it became clear that Instant Voice Clone was unlikely to produce the level of accuracy I was looking for.
My voice with accent data. This was good to test, but it didn’t help at all.
Switching to a professional voice clone
To achieve better results, I moved to a Professional Voice Clone option on the AI platform.
Unlike the instant version, this model requires significantly more training data.
- Minimum training data: 30 minutes
- Recommended training data: around 2 hours
To give the AI the best possible chance, I decided to record the full two hours. Over several months, whenever I had spare time, I recorded segments of my voice until I reached the two-hour recommendation. I didn’t use a professional microphone. Instead, I recorded everything using my laptop microphone. Surprisingly, the biggest challenge wasn’t the technology – it was simply talking to myself for two hours. To keep the recordings natural and varied, I had to speak about random topics, explain ideas, and simulate conversation as if I were speaking with another person.
It was a surprisingly tedious process. But the results made it worthwhile.
The final result: a voice that sounds like me
Once the model finished training, the difference was immediately noticeable.Â
For the first time, the AI-generated voice genuinely sounded like me.
More impressively, the model captured subtle vocal habits and mannerisms – the small details that make a voice recognisable.
After months of experimentation, the AI clone finally felt authentic.
The final voice I made. With the professional voice clone it sounds perfect.
Key lessons from creating an AI voice clone
Building a convincing AI avatar involves more than just uploading a few recordings.
Some key lessons from the process include:
- Voice cloning quality depends heavily on training data: Small datasets may work for some voices, but others require significantly more recordings.
- Instant voice cloning isn't always enough: For professional-quality results, longer recordings and higher-quality training data are often required.
- Natural speech matters: Recordings that sound like natural conversation produce better results than scripted or robotic delivery.
- Experimentation is essential: Different voices respond differently to AI models, meaning persistence and testing are often required.
Potential uses for AI avatars and voice cloning in pharma marketing
While my experiment began as a personal project, it quickly became clear that AI avatars and voice cloning have significant potential in regulated industries such as pharmaceutical marketing.
Pharma companies face a constant challenge: delivering consistent, accurate, and compliant information across multiple markets, audiences, and channels. AI avatars offer a scalable way to communicate complex information while maintaining a human presence.
Here are some potential ways the technology could be used.
Scalable medical education content
Medical education is a core part of pharmaceutical communication. AI avatars could allow organisations to create educational video content featuring a consistent presenter without requiring repeated filming sessions.
Once an avatar and voice model are trained, new content can be produced quickly by simply updating the script. This makes it easier to keep educational material up to date as clinical evidence evolves.
For example, AI avatars could be used to deliver:
- HCP educational videos
- Clinical trial summaries
- Treatment pathway explanations
- Mechanism-of-action presentations
This approach could significantly reduce production time while maintaining a consistent delivery style.
Looking for ways to enhance your eLearning and medical education provision?
Multilingual global communication
Pharmaceutical companies operate across global markets, often needing to communicate the same information in multiple languages.
AI avatars combined with voice synthesis could allow organisations to quickly generate multilingual versions of the same content without recording separate videos for every market.
Instead of filming multiple presenters, companies could create one master script and localise the voice and language while maintaining a consistent visual identity.
This could be particularly valuable for:
- Global product launch communications
- Training materials for international teams
- Educational content for healthcare professionals
Consistent brand and messaging delivery
Maintaining consistent messaging across channels is critical in pharma marketing, where compliance and accuracy are essential.
AI avatars could act as a consistent digital presenter for approved messaging. Once content has passed medical, legal, and regulatory review, the same avatar could deliver that message across multiple formats and platforms.
This could help ensure that:
- Messaging remains aligned with approved content
- Brand tone and delivery stay consistent
- Updates can be rolled out quickly when guidance changes
Training and internal communication
AI avatars could also be valuable for internal communications within pharmaceutical organisations.
Training materials, onboarding programmes, and internal updates often require repeated presentations from subject matter experts. AI avatars could help scale these communications without requiring senior experts to repeatedly record new content.
Potential use cases include:
- Compliance training
- Sales force education
- Onboarding programmes
- Internal product briefings
Faster content production
Traditional video production requires planning, filming, editing, and scheduling presenters. This process can take weeks.
AI avatars dramatically reduce this production cycle. Once the avatar and voice clone are trained, new video content can be generated much more quickly.
For pharma marketing teams operating in fast-moving therapeutic areas, this could help accelerate the delivery of educational and promotional content.
Important considerations
While AI avatars offer clear opportunities, they must be used carefully within the pharmaceutical industry.
Transparency, governance, and compliance remain essential. Any use of AI-generated presenters should be clearly disclosed, and organisations must ensure that content continues to meet medical, legal, and regulatory standards.
As with any emerging technology, the goal should not be to replace human expertise, but to augment it.
When used responsibly, AI avatars could become a valuable tool for scaling communication while maintaining accuracy and consistency.
Ready to explore AI avatars and voice clones for your business?
AI avatars and voice cloning are transforming how organisations communicate, educate, and engage with audiences. Whether you want to scale personalised content, deliver consistent messaging, or create multilingual video campaigns, Iguazu can help you bring AI-powered presentations to life.
Get in touch with our team today to discuss how AI avatars and voice clones can accelerate your content strategy and create a more engaging experience for your audience.
Frequently asked questions (FAQs)
An AI voice clone is a synthetic voice generated using artificial intelligence that replicates a real person’s voice. Platforms such as ElevenLabs analyse voice recordings to recreate tone, pitch, cadence, and speech patterns.
Some tools offer instant voice cloning using only a few minutes of audio. However, for highly realistic results, professional voice cloning models often require 30 minutes to two hours of training data.
Voice cloning accuracy depends on several factors:
- Quality of the recordings
- Amount of training data
- Consistency of speech patterns
- Background noise
- The AI model used
Some voices are simply easier for AI systems to reproduce than others.
Instant voice cloning typically uses a small amount of audio to quickly generate a synthetic voice. While fast, the results can sometimes lack accuracy.
Professional voice cloning requires significantly more training data but produces far more realistic results.
AI avatars are increasingly used for:
- Training and learning content
- Marketing and product videos
- Customer support communications
- Multilingual video generation
- Personalised business messaging
By combining an AI avatar with voice cloning, organisations can create scalable video content without needing to film new footage every time.
Yes. Many organisations are now experimenting with AI avatars trained on employees or presenters. These avatars can deliver scripted content while maintaining a consistent visual and vocal identity.
However, ethical considerations and clear consent are important when using AI to replicate a person’s likeness or voice.
Further reading
If you found this case study helpful, explore these related resources to further future-proof your pharmaceutical marketing strategy:
The Best First AI Agent in Pharma Is Probably Boring
Humanising Pharma Communication: How AI Avatars Scale Engagement and Compliance
Scale Pharma Field Coaching Compliance & Confidence with AI
10+ Innovative Web App Ideas for Pharma Marketing & Patient Support
David Pinto
Senior Project Manager
About Iguazu: We are a digital agency specialising in delivering tactical marketing solutions to the healthcare and pharmaceutical industry.
