Three Approaches to AI Music: Soundraw, Synthesizer V and MusicLM

In this February 2023 video I explore three approaches to AI music: arranging generated instrumental tracks, synthesizing vocals and research on generating music from descriptions. The products and prices I mention reflect that time.

Soundraw generates instrumental options from selected moods and genres. I choose several mood tags and preview tracks, then adjust the energy of different sections to create a buildup.

In the demonstrated pro mode, the arrangement can be adjusted by component, including melody, backing, bass, drums and fills. Tempo, duration, key and instruments offer further control over the result.

I try different genres and themes, including tutorial and technology, to show how selections produce different tracks. The interface also offers a way to find similar music and preview a track under a video.

I describe a free generation and bookmarking tier and a paid download plan then listed around $16.99 per month. Those are historical terms from 2023 and should not be assumed to be current.

Next I demonstrate Dreamtonics’ Synthesizer V, a desktop application for assigning lyrics and notes to a synthetic singing voice. I used ChatGPT to draft lyrics and chose the Natalie voice for my experiment.

I describe my own musical knowledge as limited and am not satisfied with the first result; I notice a pronunciation problem in the vocal. The example illustrates both the possibility and the need for editing.

The editor lets me place notes, enter lyrics and adjust voice parameters such as softness, breathiness, pitch and key. Moving notes and lining up syllables changes the synthesized performance.

I suggest combining an instrumental track with synthesized vocals and show the available voice options from Dreamtonics. Some voices and the fuller product version were paid offerings in the demonstration.

The final example is Google Research’s MusicLM, which generates music from text descriptions. At the time of this video Berman could only play research examples, not use it as a publicly available interactive product.

I listen to examples prompted as an arcade soundtrack and a fusion of electronic dance music with reggaeton. The examples demonstrate the research direction, but listening to a few samples is not a broad evaluation of quality or originality.

The research page also demonstrates sequences of prompts that change the music over time and conditioning on a supplied melody as well as text. I show examples of humming or whistling as melodic inputs.

I play an example connected to a visual artwork and look at instrument-specific samples. I anticipate that text-to-music tools will become more accessible, a prediction made from the research state visible in 2023.

I conclude by comparing the accessible arrangement and vocal tools with the then research-only text-to-music demonstration. Each offers a different degree of control and requires different creative input.