adlb/Audialab_EDM_Elements

Audialab - EDM Elements 2024

Introduction

This model specializes in generating high quality key-locked and tempo synced samples to support granular music production workflows. It features high musicality, robust Audio-to-Audio capabilities and is capable of generating several effects with each sample. The model can generate an infinite variety of musical components with a specialization in supersaw chord progressions, top line melody leads, bell plucks, and bass pluck riffs. All output is BPM-synced and key-locked to any note within the 12-tone chromatic scale, in both major and minor keys. Developed over several weeks, this model was trained on a custom internal dataset and features several distinct sound types:

Sound Description	Audio
1. Sine / Bell Plucks
2. Lead Square Legato Synth
3. Lead Square Warm Synth
4. Bass Plucks
5. Supersaw Leads & Chord Progressions

Furthermore, the model is capable of generating multiple post-processing effects, with several different levels of control based on the prompt alone. All effects can be prompted individually or together.

1. Reverb - Small Reverb, Medium Reverb, High Reverb
2. EQ Sweeps - Rising Low-Pass, Falling High-Cut
3. Gate Effect - Quarter-Beat gate, Half-Beat Gate

The model has high Audio-To-Audio / Style Transfer capabilities. For example you can upload a .wav sample and have the model turn the sample into say a Supersaw version with a rising low-pass.

Prompt	Settings	Before	After
Supersaw, high reverb, rising low-pass	CFG: 10.0 / Noise: 1.29

Model Features

Multiple Types of Sample Generation: - Outputs multiple types of high quality EDM-Centric samples with several variations on melody and chord progressions.
Dynamic FX Chain: - Add multiple effects based on prompt alone (prompts above) - i.e add "Medium Reverb" for a medium reverb effect on the output.
Tonal Versatility: - Generates stems in any key across the 12-tone chromatic scale, in both major and minor scales.
Triplet Time Signature Support: - Add "triplets" to any prompt and get triplet outputs.
AI Style Trasnfer / Audio to Audio: - The model has robust AI Style Transfer capabilities.
Speed Controls Independent of BPM: - The model can be prompted with "Slow Speed", "Medium Speed" & "Fast Speed" that is independent of BPM. i.e. ask for an arp at a "Fast speed" at 140BPM vs "slow speed" at 140BPM and the model will subdivide the notes accordingly while staying within the same bpm.
Simplified Scale Notation: - Scales are written using sharps only in the following format:

Minor Scales
A minor, A# minor, B minor, C minor, C# minor, D minor, D# minor,
E minor, F minor, F# minor, G minor, G# minor

Major Scales
A major, A# major, B major, C major, C# major, D major, D# major,
E major, F major, F# major, G major, G# major

Training Methodology

This model was designed to understand and generate several types of samples:

Chord Progression Samples
Chord Progression with Melody Samples
Melody Only Samples
Supersaw Dance Chord Progressions
Bass Plucks / Riffs
Arps

By exposing the model to various musical motifs and distinct sample differences, it has SOTA musicality and melody structure with output samples ready to use in music production.

Prompt Structure

Both the Audialab Engine (VST) and RC Github fork are best used when working with this model. The RC Github also features a randomized prompt button tuned to this models metadata.

If you wish to prompt outside of these interfaces - then to ensure the best results use the following format for your prompts:


[Sound Type], [Modifier][Chord Progression], [Modifier][Melody Type], [Key], [FX], [BPM], [Bar Count]

[Sound Type Prompts]

1. Bell Plucks: - 'Pluck, Sine, Bright, Clean, Bell'
2. Lead Square Legato Synth - 'Lead, Square, Synth, Buzzy, Legato'
3. Lead Square Warm Synth - 'Lead, Saw, Synth, Warm, Supersaw'
4. Pluck Bass - 'Bass, Punchy, Pluck, Clean, Sub, Sine'
5. Supersaw Leads & Chord Progressions 'Supersaw, Synth, Warm, Saw'

Major prompt terms and their effects

Chord Progression - produces chord progressions
Arp - produces arps
Melody - adds melodies
Triplets - outputs triplet time
Bounce - more syncopassion/ off beat rhythm
Epic - encourages more interesting melodic output
Simple - slower / simpler melodies / chord progressions
Slow speed - slower chord progressions, arps and melodies
Medium speed - encouraged 4/4 chord changes or melodies
Fast speed - higher subdivisions / faster arps and melodies
Complex - adds complex melodies or chord structure.
Rising - encourages rising arps / melodies
Falling - encourages falling arps / melodies
See random prompt document for further terms / examples HERE

Prompt Examples with Audio

Prompt	Example 1	Example 2	Example 3
Bell pluck, chord progression, top catchy melody, E major, high reverb, 8 bars, 100BPM
Saw, supersaw, chord progression, top melody, E minor, rising low-pass, 4 bars, 128BPM
Sine, Bass, Bounce, Catchy Melody, G minor, 8 bars, 110BPM
Sine, Bell, Pluck, chord progression with slow melody, B minor, high reverb, rising low-pass, 8 bars, 140BPM
Supersaw, dance chord progression, top triplet melody, F# minor, medium reverb, 8 bars, 130BPM
Synth, saw, fast speed, falling arp, medium reverb, falling high-cut, C major, 4 bars, 150 BPM
legato, square, catchy melody, high reverb, E minor, 8 bars, 130PM
supersaw, complex, epic, arp, D# minor, 8 bars, 140BPM

Audio to Audio / Style Transfer Showcase

Prompt	Settings	Before	After
supersaw, medium reverb	CFG: 9.0 / Noise: 1.16
Pluck Bell, Medium Reverb	CFG: 8.2 / Noise: 1.21
Supersaw, high reverb, rising low-pass	CFG: 10 / Noise: 1.29

Prompt Examples

You can find a handy list of starter prompts HERE

BPMs/Bars:

The BPMs ranged from as low as 100BPM up to 150BPM. The main denominations are 100BPM, 110BPM, 120BPM, 128BPM, 130BPM, 140BPM, 150BPM.

There are 2 bar settings: 4 bars and 8 bars.

Useage Guide

VST Support

This model has direct VST compatibility in Audialab Engine.

Gradio Interfaces

You can find a direct link to a custom GitHub interface here.

If you wish to use the original Stable Audio Github then you can follow this link.

You will find 2 checkpoints files in the repo:
Audialab_EDM_Elements.ckpt - Full 32 bit model
and
Audialab_EDM_Elements_Small.ckpt - 16-bit low VRAM version
along with the associated config file - model_config.json

To use the model you would choose either the full model or the 16-bit quantized model. Once chosen place the .ckpt file and the config .json inside their own sub-folder in the "models" folder and launch the gradio.

Dataset Breakdown

Overview

Total .wav files: 606,976
Total Size: 58.03 GB (pre-encode)
Sample Rate: 44100 Hz
Sample Breakdown

Sound Type	Sample Count
Bell Pluck	100,553
Warm Supersaw	100,320
Square Lead	54,475
Square Buzzy Lead	54,845
Pluck Bass	54,732

Augment Breakdown

Augment Category	File Count
Low Cut Rise w/ Half Gate	22,763
High Cut Faller w/ Half Gate	22,849
Reverb Augments	90,180
Gate Augments	30,045
High Cut Faller Quarter Bar Gate	22,703
Low Cut Rise w/ Quarter Gate	22,589
High Cut & Low Cut Augments	30,922

While the dataset appears quite large it featured heavy augmentation. For example a single melody could be stretched into multiple keys and multiple tempos with slight changes for variance. This means the training time must be carefully watched to ensure it doesn't overfit to specific melodic patterns.

Further the FX augments used the original stem but with only say a "high reverb" preset applied. This stretches the data further and lets the model hone in one this specific component of the sound (as it's seen both no-FX versions along with identical FX applied versions)

Technical Specifications

Platform: Runpod
Monitoring Tool: Weights and Biases
Total Steps: 33,436
Final Checkpoint Step: 28,452
Learning Rate: 5e-5
Dataset type: Pre-encode
Optimizer: AdamW
Scheduler: InverseLR
Batch Size: 32
Hardware: 2x NVIDIA A40 GPUs

See config file for further details.

Limitations and Biases

The model itself has very high musicality and can essentialy be used as an infinite sample + melody generator. A limitation however of the model is since it was created using our own dataset it is not set up for genre prompting and/or sound types outside of the ones mentioned above. i.e. A prompt like "House, Pad, Guitar" would not produce such samples.

In the future with scale - genre prompting and additional sound types can be released in future models.

Lastly as mentioned it has very robust audio-to-audio capabilities however the input audio should be within a similiar frequency space as the desired output. For example if you had a piano progression and wanted to change it to a supersaw then it will do very well. But if you put a bass sample as the input (which usually falls within the lower frequency spectrum) and ask for a supersaw version (which falls within the mid-to high frequency) then the model will struggle with the conversion. A workaround is to pitch the input sample up or down to get near to the desired output frequency range. i.e. if you wanted to AI style transfer a piano melody to a bass pluck then pitching the input piano sample down a few semitones and then converting to bass will have better results.

Additional details

Dataset: Internal IP + Tooling
Gate Module: Custom NumPy FX
Reverb Effect Solaris Reverb
EQ: Simple EQ Effect (riser/faller)

License

This model is licensed under the Stability AI Community License. It is available for non-commercial use or limited commercial use by entities with annual revenues below USD $1M. For revenues exceeding USD $1M, please refer to the LICENSE for detailed terms.

adlb
/

Audialab_EDM_Elements