Artificial intelligence did not begin with the public rise of tools such as ChatGPT. Audio researchers and developers had already been exploring machine learning for years. At Bitsonic, I used a deliberately narrow system: instead of trying to solve general tasks, I trained it to study and generate specific bass-guitar material.
Building a focused training system
The goal was to generate new monophonic timbres from a dedicated training set. The resulting sounds became part of AI Bass Ultimate, but reaching that point required both a deep-learning algorithm and a carefully prepared collection of bass-guitar samples.
The obvious question is why generated bass sounds are useful when high-quality recordings already exist. For this early experiment, the value was not replacing a recorded instrument. It was testing whether a focused model could learn useful characteristics from the source material and produce variations that were not simple copies of individual samples.
From recordings to generated samples
I grouped the source recordings by pitch, playing style and tonal character. The files were stored at 24-bit/88.2 kHz. The 24-bit format provided comfortable headroom during preparation, while the higher sample rate preserved source detail even though it was not decisive to the final result. The categorized sounds were then converted into high-resolution spectrograms that the system could process.
Monophonic recordings made the experiment more manageable. A plucked bass note contains a fundamental frequency and a series of harmonics that remain easier to distinguish on a spectrogram than overlapping notes in complex polyphonic material.
After analyzing the training material, the system generated requested pitches as spectrograms, which were then converted back into audio. I reviewed the results and fed the more useful examples into later training rounds. This was a narrower approach than OpenAI's 2020 Jukebox research, which modeled much more complex raw-audio music. In simple terms, our experiment generated a visual frequency representation first and resynthesized it as sound afterward.
Where the process can go next
During the experiment, the system became better at separating useful material from noisier examples and at producing variations from a smaller training set. Morphing also made it possible to combine characteristics of two source sounds. These results were promising, but they did not prove that a few samples could automatically produce a complete, production-ready instrument.
The harder challenge is not generating a different waveform; it is creating a sound that is both genuinely useful and meaningfully different from results already possible with synthesis, sampling and conventional processing. That remains the most interesting part of the experiment. The broader effect of these tools on working musicians is explored in The Impact of Artificial Intelligence on Composers.