Teaching Generative Audio Models Non-Western Rhythms: Fine-Tuning ACE-Step 1.5 with LoRA on 6/8 South Indian Percussion

Boobalan Arjunan — Fri, 16 Oct 2026 10:55:15 +0000

State-of-the-art music generation models like ACE-Step 1.5 are technical marvels, but they carry a known systemic blindspot: they were trained predominantly on Western 4/4 structures.

When you prompt foundation models for traditional South Indian acoustic percussion like Parai, Thappu, or high-energy Dappankuthu, the output falls flat. The reason is musical and mathematical: our folk drumming relies on 6/8 compound time signatures (tha-ki-tta, tha-ki-tta) and rapid acoustic transient spikes that standard cross-attention layers fail to resolve on standard binary grids.

To solve this, I fine-tuned and open-sourced SouthIndiaParaiLoRA on top of the ACE-Step 1.5 foundation architecture.

🔬 Engineering & Training Setup

Base Architecture: ACE-Step 1.5 (acestep-v15-x1-base & turbo compatible)
Dataset: 393 studio-quality stems of authentic Parai, Thappu, Thappattam, and Urumi loops (80 – 172 BPM)
Hardware: 30 epochs on an NVIDIA A40 (48GB VRAM) via RunPod
LoRA Configuration: Rank 64, Alpha 128
Optimization: Slashed final loss by 75.5% (from a 1.97 baseline down to 0.48)

🛠️ How to Run in ACE-Step Playground & ComfyUI

The model weights are released in standard .safetensors format:

1. ACE-Step Playground

Clone the engine: ACE-Step-1.5-PlayGround--Engine
Drop adapter_model.safetensors and adapter_config.json into:

   checkpoints/SouthIndiaParaiLoRA/final/

Launch the WebUI and select SouthIndiaParaiLoRA from the LoRA dropdown.

2. ComfyUI Workflow

Load adapter_model.safetensors directly into your standard ACE-Step 1.5 LoRA loader node.

🎯 Example Prompt

Prompt: High energy South Indian dappankuthu street festival drumming, massive thappu ensemble, authentic 6/8 compound folk rhythm at 145 BPM, urumi and parai dynamic groove
Lyrics: [Instrumental]
LoRA Scale: 0.85

🎧 Model Weights & 7 Audio Demos

The adapter weights and 7 pre-rendered audio demos across different tempos and rhythmic variations are available under Apache 2.0 on Hugging Face:

👉 [Hugging Face Model Card & Demos]

(https://huggingface.co/boobalanit/ACE-Step-1.5-SouthIndiaParaiLoRA)

I'd love to hear feedback from fellow researchers, audio engineers, and developers experimenting with ACE-Step or regional music synthesis!

WaveWash: Clean, Repair, and Dedupe Audio Datasets Before ML Training

Boobalan Arjunan — Wed, 14 Oct 2026 11:09:26 +0000

Training speech models (Whisper, TTS, voice cloning) often breaks down because of dirty audio data: clipping, DC offset, background noise, low SNR, and duplicate samples quietly inflating validation metrics.

I built WaveWash to automate audio quality control and data prep so developers don't have to write one-off scripts or manually listen to thousands of audio files.

What WaveWash Does

Tuned Presets: Built-in profiles for ASR (Whisper), TTS, Speaker ID, and Music Generation.
Safe Auto-Repairs: Corrects DC offset, mild clipping, and normalizes loudness (EBU R128).
Automated Segregation:
- clean/ : Audio that passed all checks.
- _error/ : Rejected files categorized by reason (low_snr/, clipping/, etc.).
- _duplicates/ : Exact and near-duplicates set aside (nothing is deleted).
- _logs/ : Complete audit logs (CSV, JSON, and Markdown summaries).
Zero-Config Local Web UI: Ships directly inside the Python package — no Node.js build required.
Non-Destructive: Your original files are never modified.

Quick Start

Install and launch the local web UI in two commands:


bash
pip install wavewash
wavewash app

It opens your browser at http://127.0.0.1:8420.

You can also run it directly from the terminal:

wavewash run --input ./raw_audio --output ./clean_audio --purpose asr


Links
PyPI: https://pypi.org/project/wavewash/
GitHub: https://github.com/BoobalanArjunan/wavewash
It's currently in early alpha (v0.1.0a1). Feedback, issues, and feature suggestions are very welcome!

DEV Community: Boobalan Arjunan