Concert crowd

Late South Korean folk singer stages AI ‘comeback’

Image credit: Dreamstime

A South Korean audio company has used its AI technology to recreate the voice of a native folk-rock singer who has been dead for nearly 25 years.

Developed by AI company Supertone, the voice AI system 'Singing Voice Synthesis' (SVS) learned 20 songs of the popular South Korean singer Kim Kwang-seok, based on a training tool with over 700 Korean songs.

The SVS technology learns voices by listening to multiple songs with corresponding notes and lyrics, the firm said.

To improve the accuracy of the late singer’s voice, the selection of songs learned within the training tool allowed the system to mimic a new song in Kim’s own style, according to Supertone co-founder and chief operating officer Choi Hee-doo. 

To recreate these voices, Supertone said it first acquired voice samples for the raw material for its training AI models. These samples were then enhanced by the company’s proprietary source separation technology. 

Using this elaborately processed training data, Supertone then trained a neural network using a high-performance computing server. The company claims its voice synthesis and cloning neural networks boast the most hyper-realistic quality for audio.

As part of a television show to be aired today (29 January) by South Korean broadcaster SBS, 'Competition of the Century: AI vs Human', the late singer’s voice will be used to perform 'I miss you' - a ballad released in 2002 by the country’s famous singer Kim Bum-soo.

“What we were most scared and worried about was asking Kim’s family for their permission,” said Kim Min-ji, a producer of the show. “When we played the AI audio to his family, they were very happy with it. They were wary at first, but after listening to the outcome, they said it was like Kim Kwang-seok came back alive.”

To sound accurate, it was important that the voice and performance carried enough feeling, said Nam Sang-moon, chief producer for the show. “It’s about singing a song with emotions. Each note should be connected naturally, hence emotions can be delivered, which is the key point of the technology. The developers trained the AI to exactly imitate human’s vocal organs.”

Supertone is a start-up founded in 2020 that provides AI audio solutions for content creators, with Kim Kwang-seok’s voice becoming the latest in their repertoire. According to the company, it is the world’s first voice synthesis solution for singing and acting. 

Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.

Recent articles