APINAJA/inverse-fx-phase1-effect-type

Inverse FX Phase 1 — Effect-Type Dataset

Paired dry / wet audio for multi-label effect-type prediction (Inverse FX Chain Predictor, Phase 1). Generated from APINAJA/inverse-fx-source-audio-cleaned with Pedalboard (pedalboard 0.9.25).

• Effects: EQ, Compressor, Reverb, always applied in the fixed order EQ -> Compressor -> Reverb
• Label: binary vector [EQ, Compressor, Reverb] (effect presence only)
• Sources: 60 dry clips (Bass: 20, Drums: 20, Mixture: 20)
• Pairs: 420 (7 combinations per dry clip)
• Audio: 44100 Hz, 1 channel(s), 10.0 s clips, 32-bit float WAV
• Effect parameters: random, reproducible with seed 42

Effect combinations

#	effect_chain	label
1	EQ	[1, 0, 0]
2	Compressor	[0, 1, 0]
3	Reverb	[0, 0, 1]
4	EQ + Compressor	[1, 1, 0]
5	EQ + Reverb	[1, 0, 1]
6	Compressor + Reverb	[0, 1, 1]
7	EQ + Compressor + Reverb	[1, 1, 1]

Splits (song level, inherited from the source dataset)

split	songs	dry clips	dry/wet pairs
train	14 (70%)	42	294
validation	3 (15%)	9	63
test	3 (15%)	9	63

Naming

field	example
source_sample_id	1-Bass
sample_id	1-Bass-EQ-Compressor
source_id	1
source_category	Bass
effect_chain	["EQ","Compressor"]
label	[1,1,0]

Files

• data/<split>-*.parquet — one row per pair, dry_audio and wet_audio as playable Audio (the struct's path keeps the original WAV path); all other columns as in metadata/<split>.csv
• data/<split>/dry/<source_sample_id>.wav — dry clip (shared by its 7 wet versions)
• data/<split>/wet/<sample_id>.wav — processed audio
• metadata/<split>.csv — one row per pair: ids, paths (dry_audio, wet_audio), effect_chain, label, eq_label / compressor_label / reverb_label, exact parameters per effect (eq_params, compressor_params, reverb_params, JSON; empty when the effect is absent), normalised parameters, per-effect signal change and the per-sample seed
• metadata/sources.csv — how each dry clip was cut and normalised
• metadata/dataset_summary.json, metadata/config_used.yaml — full generation settings

Usage

With 🤗 Datasets (dry_audio / wet_audio decode to arrays):

import json
from datasets import load_dataset

ds = load_dataset("APINAJA/inverse-fx-phase1-effect-type")
row = ds["train"][0]
row["dry_audio"], row["wet_audio"]     # Audio
label = json.loads(row["label"])       # [EQ, Compressor, Reverb]
params = json.loads(row["parameters"])

Or directly from the WAV files and CSV metadata:

import json, pandas as pd, soundfile as sf
from huggingface_hub import snapshot_download

root = snapshot_download("APINAJA/inverse-fx-phase1-effect-type", repo_type="dataset")
train = pd.read_csv(f"{root}/metadata/train.csv")
row = train.iloc[0]
dry, sr = sf.read(f"{root}/{row.dry_audio}")
wet, _ = sf.read(f"{root}/{row.wet_audio}")
label = json.loads(row.label)          # [EQ, Compressor, Reverb]
params = json.loads(row.parameters)    # {"EQ": {"frequency": ..., "gain": ..., "q": ...}, ...}

License

Derived from MUSDB18 — non-commercial research use only, under the original dataset terms.