Cloning my father’s voice from 2000 and the transformative promise of neural networks on Parkinson’s Disease research.

In May 2000, my family moved from Omaha, NE to a small town in Southwest Missouri called Nixa. I was seven and my brother was two.

Six years later, my father was diagnosed with Parkinson’s Disease (PD) after initially going to see the doctor about a slight tremor he had on one side.

He went two days without telling anyone. Even Mom.

Let me start by saying my father is alive. I am lucky to still have both of my parents in my life. I just saw them last weekend for Mother’s Day and my brother’s birthday. I am so thankful for that time.

But he has changed. His voice has changed as happens with many Parkinson’s patients. Many living with PD exhibit a muted vocal expression, speaking in a monotonous tone and conveying limited emotional range. But it’s been a slow burn — thank goodness.

We moved for work, which was the very thing that drove my father to get out of bed every morning. And when this new, young hotshot school Superintendent came in from Omaha, the local news took note.

The first time my brother saw this footage was in 2017 at my father’s retirement roast. I was sitting right next to him. The look on his face as he watched and listened has stuck with me over the years.

He doesn’t remember my dad’s voice sounding like that.

He was excited and also, I think, a little bit disappointed that he didn’t get more time with this more charismatic-seeming version of our father.

AI will never replace that, of course, but in the whirlwind of dystopic AI predictions, there does seem to be promise for good as well.

Bank Your Voice

It’s shockingly easy to clone voices now with just a few seconds of audio.

Using GarageBand to cut my father’s audio from the KY3 interview.

When I heard about ElevenLabs’ VoiceLab, a voice cloning tool, this idea immediately came to mind. I isolated my father’s voice from the interview and then simply uploaded that audio export (.mp3) to ElevenLabs and bam.

It was my father’s voice from 2000, saying anything I typed into a box.

Obviously there are a ton of ethical questions here, which is why you should always ensure you’ve obtained explicit and informed permission before using someone’s voice. But I was able to create an audio book version of his children’s book, Golf Ball Envy, in his former voice. Sharing this with my family brought about laughter and genuine interest.

I asked him to read a short note to my brother, Jacob.

https://soundcloud.com/elijah-kleinsmith-903587942/real-dad-message-to-jacob

Here’s the 2000 version of him reading it.

https://soundcloud.com/elijah-kleinsmith-903587942/ai-dad-message-to-jacob

It sort of permanently sounds like he’s giving an interview to a TV station to me, but given better audio quality inputs, you can make it sound identical.

Like in this sample I created using audio from an iPhone video shot in an empty room. It’s my brother’s real voice as he records this for Mother’s Day.

https://soundcloud.com/elijah-kleinsmith-903587942/real-jacob-short

Now here’s the AI version of him saying hi to my girlfriend, Julia.

https://soundcloud.com/elijah-kleinsmith-903587942/ai-jacob-short

It even allows you to hear what your voice would sound like in another language, as though you spoke it fluently. Here’s Jacob in Japanese.

https://soundcloud.com/elijah-kleinsmith-903587942/ai-jacob-japanese

I’m not the first person to think about this. Intel has a new project called I Will Always Be Me which “banks” the voices of those affected by motor neurone diseases so that they can communicate using their own voice.

More “AI for Good” News

Neural networks, like the voice cloning tool I used here, show real promise of predicting Parkinson’s Disease before symptoms show up.

I uploaded the full white paper to ChatPDF.com, which uses the power of GPT to allow users to ask PDF files questions about its content.

Me: What is difference between the neural network models and the traditional machine learning models?

PDF: Neural network (NN) models differ from traditional machine learning (ML) models in their ability to learn complex non-linear relationships between input features and output variables. NN models are composed of multiple layers of interconnected nodes that can learn and extract features from raw data without the need for manual feature engineering. In contrast, traditional ML models typically require manual feature selection and engineering to extract relevant information from raw data. NN models can also handle high-dimensional data with many correlated features, which can be challenging for traditional ML models. Additionally, NN models can provide interpretable results by identifying significant biomarkers that contribute to disease prediction, which is important for clinical applications. Overall, NN models offer a powerful tool for analyzing large metabolomics data sets and developing accurate prediction models for disease diagnosis.

Me: Does it mention commercial availability of testing?

PDF: No, the file does not mention anything about the commercial availability of testing. It mainly focuses on the development of an interpretable neural network (NN) framework to accurately predict Parkinson’s disease (PD) from blood plasma metabolomics data using machine learning (ML) approaches. The study aims to provide opportunities for early diagnosis of disease by analyzing metabolite biomarkers and to develop accurate prediction models for disease diagnosis using large metabolomics data sets.