AI technology is reaching a whole new level. Some Facebook engineers have created voice clones of Bill Gates. This is as computer-generated speeches become more achievable.

 

In the clips, you can listen to what sounds like Bill Gates speaking some phrases. However, a machine learning system generated each of the voice clips. Engineers at the social media giant company, Facebook, created and designed the MelNet.

 

Other voices have also been mimicked. Like that of renowned scientist, Stephen Hawking and British anthropologist, Jane Goodall.

 

Listen to the embedded clips below to hear the generated Bill Gates voice.

 

“Two plus seven is less than ten.”

 

“A cramp is no small danger on a swim.”

 

 

 

Bill Gates did not exactly say those words. A 452-hour dataset of TED talks was used to train the MelNet. They also trained it using data from audiobooks because of the animated manner in which the speakers spoke.

 

The quality of the voice clone has seen improvements over the years. A lot of progress was made in 2016 with the unveiling of the WaveNet and Sample RNN. The WaveNet is a machine learning text-to-speech program. Google’s AI lab, DeepMind, which powers the Google Assistant, created it.

 

WaveNet and SmapleRNN had the basic approach of feeding the AI system with data used to analyse the human voice. But Facebook’s more advanced MelNet ‘uses a richer and more informationally dense format to learn to speak: the spectrogram.’

 

 

MelNet

Facebook’s researchers noted that the MelNet was superior at capturing high-level structure. It can be more detailed in learning the consistencies of a speaker’s voice. This is because the data captured in the spectrogram is more compact than that in audio waveforms. The density then allows the algorithm to produce more consistent voices.

 

However, there are some limitations. The MelNet is unable to replicate how the human voice can change over a long period of time. It is unable to sustain the changes that occur with tension in the voice when reading a page of text.

 

MelNet can also incredibly generate not just realistic voices but music too. Even though the music bit could use some work, it is still rather impressive.

 

Meanwhile, this new technology might be problematic in the sense that people can be quoted to say something they didn’t. The same ways doctored images still cause major problems on the tabloids.