You can deploy this model and start earning money today!

rikhoffbauer2/
dj-reproducer

Name: dj-reproducer
Availability: InStock

ml-intern

2008.10267

Views

No views yet

DJ Set Reproducer

A 5-stage audio analysis pipeline that reverse-engineers DJ sets into accurate tracklists with timestamps and transition analysis.

Repository: https://huggingface.co/rikhoffbauer2/dj-reproducer

🚀 Gradio UI

This repository includes a Gradio app (app.py) for interactive use. To run it:

bash

pip install -r requirements.txt
python app.py

Or deploy as a Hugging Face Space:

Go to https://huggingface.co/new-space
Select Gradio SDK
Clone this repo's files into it
Add system packages: ffmpeg, libchromaprint-tools, libsndfile1

Pipeline Overview

Stage	Purpose	Output
1	Fingerprint DJ set every ~30s, identify songs via AcoustID	`stage1_fingerprints.json`
2	Download identified songs from YouTube, confirm identity	`downloads/` + `stage2_confirmed.json`
3	Align original songs with DJ set to find which part is used	`stage3_alignments.json`
4	Detect transitions (cue-out → cue-in) using DTW path analysis	`stage4_transitions.json`
5	Analyze transition techniques (cut, crossfade, EQ, etc.)	`stage5_analysis.json`
—	Generate final tracklist	`tracklist_final.json` + `tracklist.txt`

Installation

bash

# Clone repository
git clone https://huggingface.co/rikhoffbauer2/dj-reproducer
cd dj-reproducer

# System dependencies (Ubuntu/Debian)
sudo apt-get install -y libchromaprint-tools ffmpeg libsndfile1

# macOS
brew install chromaprint ffmpeg

# Python dependencies
pip install -r requirements.txt

Prerequisites

AcoustID API Key: Register free at https://acoustid.org/login
ffmpeg and fpcalc (chromaprint tools) installed
Python 3.9+

Usage

Full Pipeline (CLI)

bash

python -m dj_reproducer.cli process \
    --input "my_dj_set.mp3" \
    --output-dir ./results \
    --acoustid-key YOUR_ACOUSTID_KEY

Individual Stages (CLI)

bash

python -m dj_reproducer.cli stage1 --input set.mp3 --acoustid-key KEY
python -m dj_reproducer.cli stage2 --tracklist results/stage1_tracklist_raw.json --acoustid-key KEY
python -m dj_reproducer.cli stage3 --input set.mp3 --confirmed results/stage2_confirmed.json
python -m dj_reproducer.cli stage4 --input set.mp3 --alignments results/stage3_alignments.json
python -m dj_reproducer.cli stage5 --input set.mp3 --transitions results/stage4_transitions.json --alignments results/stage3_alignments.json

Gradio UI

python app.py

Then open http://localhost:7860 in your browser.

Algorithm Details

Stage 1: Fingerprinting

Window strategy: 30s windows with 25s hop (5s overlap) for redundancy
Pipeline: ffmpeg slice → fpcalc chromaprint → AcoustID lookup
Rate limiting: 0.35s delay between requests (AcoustID free tier: 3 req/s)
Consolidation: Merge overlapping windows of same track into contiguous segments

Stage 3: Alignment (ISMIR 2020 validated)

Features: Beat-synchronous CENS chroma
Coarse alignment: Chroma cross-correlation (O(n) speed)
Refinement: Subsequence DTW (librosa.sequence.dtw(subseq=True)) on ±30s window
Key invariance: All 12 circular chroma shifts tried
Validation: Match rate ≥0.4 threshold from Serra et al. ISMIR 2020

Stage 4: Transition Detection

Method: 32-beat rolling window of diagonal DTW moves
cue-out: Last beat where preceding 32 beats are mostly diagonal (>0.8)
cue-in: First beat where succeeding 32 beats are mostly diagonal
Insight: DJs cluster transitions at 32-beat phrase boundaries

Stage 5: Technique Classification

crossfade: RMS slopes show outgoing decreasing + incoming increasing
cut: Abrupt RMS jump (>0.5) with low correlation to sum of originals
eq_filter: Spectral centroid drops significantly during transition
overlay: High correlation between mix and sum of original tracks

Data Contracts

Each stage reads/writes JSON files:

File	Schema
`stage1_fingerprints.json`	`{timestamp: {offset_seconds, duration_seconds, fingerprint, acoustid_results}}`
`stage1_tracklist_raw.json`	`{track_id: {title, artist, mbid, start_seconds, end_seconds, confidence}}`
`stage2_confirmed.json`	`{track_id: {title, artist, youtube_url, download_path, confirmed}}`
`stage3_alignments.json`	`{track_id: {offset_seconds, confidence, key_shift_semitones, beat_frames_mix, beat_frames_ref}}`
`stage4_transitions.json`	`[{track_id, title, artist, type: cue_out/cue_in, timestamp_seconds, diagonal_ratio}]`
`stage5_analysis.json`	`[{timestamp_seconds, outgoing_track, incoming_track, technique, confidence, ...}]`
`tracklist_final.json`	`{tracks: [{number, artist, title, start, end, start_seconds, end_seconds, key_shift, transition_technique}]}`
`tracklist.txt`	Human-readable plain text tracklist

Project Structure

dj_reproducer/
├── __init__.py
├── stage1_fingerprint.py   # Windowed fpcalc + AcoustID lookup
├── stage2_acquire.py     # yt-dlp download + confirmation
├── stage3_align.py       # Beat-sync CENS chroma + subsequence DTW
├── stage4_transitions.py # 32-beat rolling diagonal detection
├── stage5_analyze.py     # RMS/spectral comparison for technique classification
└── cli.py                # Orchestration CLI
app.py                      # Gradio UI
requirements.txt
README.md

References

Serra et al., "Unmixing DJ Sets with Differential Phase Analysis", ISMIR 2020 (arxiv:2008.10267)
- 80.4% cue-in hit rate within 30s, 92.9% within 60s on 20,765 real DJ transitions
- Validates beat-synchronous CENS chroma + subsequence DTW approach
AcoustID API: https://acoustid.org/webservice
Chromaprint: https://github.com/acoustid/chromaprint
librosa DTW: https://librosa.org/doc/latest/generated/librosa.sequence.dtw.html

License

MIT License — See repository for details.

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

Try ML Intern: https://smolagents-ml-intern.hf.space
Source code: https://github.com/huggingface/ml-intern