The Weirding Module – A Robot Mouth Sings

With the recent release (and removal) of several fake Drake tracks that claim to be "generated" by AI. It's important to note that these tracks, their lyrics and their music are written by humans and that the only AI element is the cloned voice of the singer which, though impressive in its simulation of Drake's voice, is basically either an effect applied to a real singer or a vocal synth that needs to be played and edited to achieve any level of rhythm or expression. This hype is the latest in a string of false claims about AI music and artists including Instagram influencer Miquela, CGI rapper FN Meka and Chinese company Tencent Entertainment's slew of virtual stars none of which use generation as we might think of it when we think of image generation.

Despite all the doom-saying current AI music tech is far from fully automated. Unlike AI image generation, generating music from prompts is in its infancy, and results from Google's MusicLM (for instance) are very rough. Calling music that uses vocal synths taught to imitate specific singers "generated" is a seemingly deliberate deception when the process to make them is almost the same as any other studio production. Perhaps it's a desire for music to work this way, for great music to be able to be produced in seconds without spending years practicing instruments, using technology or learning theory but it comes across as an attempt to devalue the work of musicians to the advantage of big tech companies. So as this technology grows and evolves it's important to analyze each claim about AI to differentiate the real capabilities from the hype. In this spirit I present a "A Robot Mouth Sings" a mix that tunes in to the reality of voice synths both AI and non-AI throughout the years.

Mechanically replicating the human singing voice goes back at least as far as the 17th Century and this episode opens with a Bach piece using the Vox Humana stop on the church organ, from there we move to Tomita's imitation of Beatles songs on his Moog, followed by various voice presets and general midi sounds from the Roland D-50 and Casio and then into some music using the revolutionary Fairlight CMI which combined both sampling and synthesis.

We listen in to AI experiments from Mouse on Mars, Jlin and Holly Herndon, Google MusicLM (real prompt based AI generated music) and artist Carrick Bell before exploring some super recent Tencent Entertainment's tracks, a Chinese corporation creating virtual pop stars with their AI voice cloning too Lingyin engine. We focus on the virtual star Beyond AI a voice clone of the real-life singer Yang Chaoyue, the two singers meet in the Beyond AI's debut song and video "Metaverse" and Beyond AI also features on last month's track 我们的时代 (Our Time) featuring 15 virtual singers created using Lingyin engine. It should be noted that far from claiming these tracks to be generated by AI the team that writes the music, lyrics and does the programming is extensively credited forming another selling point of the music.

While these developments in China are super recent virtual pop stars are nothing new in Japan. The Yamaha PLG-SG 100 was the earliest dedicated vocal synth (made in the 90s and copied in the electron monomachine) by the mid 2000s it had developed into Vocaloid – a piece of software which shipped not only with voice presets but with avatars for each voice. The community that formed around the singers, composers and fans is still thriving after more that 15 years. The episode rounds out with some recent, unhinged Vocaloid tracks featuring the most famous virtual singer – Hatsune Miku.