README.md

---
license: cc-by-4.0
task_categories:
- audio-classification
tags:
- audio
- synthesizer
- audio-effects
- parameter-estimation
- synthetic
pretty_name: StemFlipper synth/effects parameter-estimation scaffold
configs:
- config_name: synth
  data_files:
  - split: train
    path: synth/train-*
- config_name: effects
  data_files:
  - split: train
    path: effects/train-*
---

# StemFlipper — synth & effects parameter-estimation dataset (scaffold)

A synthetic **(audio → parameters)** dataset for the inverse problems StemFlipper
targets: recover a synth patch from its sound, and recover an effect chain from
wet audio. No public dataset pairs real audio with the synth-patch / effect-chain
parameters that produced it — this fills that gap with deterministic synthetic
generation. **The moat is the generator + seeds, not stored audio**: every example
here regenerates bit-for-bit from its seed.

## Configs

- **`synth`** — torchsynth `Voice` renders `(audio, params)`; `params` is the
  78-dim normalized parameter vector (`adsr_1.attack`, `vco_1.tuning`, …).
- **`effects`** — a known `dasp-pytorch` chain (`parametric_eq, compressor, distortion`)
  applied to clean synth voices → `(wet audio, params)`; `param_names` labels each
  value.

```python
from datasets import load_dataset
synth = load_dataset("nakas/stemflipper-dataset", "synth", split="train")
fx    = load_dataset("nakas/stemflipper-dataset", "effects", split="train")
```

## Reproduce / extend from seeds

The published splits are a small demo. Regenerate or scale up deterministically:

```python
from dataset.synth_gen import SynthGenConfig, iter_examples
list(iter_examples(SynthGenConfig(batch_indices=[0, 1, 2])))  # 3 batches
```

Generation spec (seeds):

```json
{
  "sample_rate": 44100,
  "synth": {
    "generator": "dataset/synth_gen.py",
    "batch_size": 32,
    "batch_indices": [
      0,
      1
    ],
    "param_names_count": 78
  },
  "effects": {
    "generator": "dataset/effects_gen.py",
    "chain": [
      "parametric_eq",
      "compressor",
      "distortion"
    ],
    "seeds": [
      0,
      1,
      2,
      3
    ],
    "n_dry": 8,
    "param_names": [
      "parametric_eq.low_shelf_gain_db",
      "parametric_eq.low_shelf_cutoff_freq",
      "parametric_eq.low_shelf_q_factor",
      "parametric_eq.band0_gain_db",
      "parametric_eq.band0_cutoff_freq",
      "parametric_eq.band0_q_factor",
      "parametric_eq.band1_gain_db",
      "parametric_eq.band1_cutoff_freq",
      "parametric_eq.band1_q_factor",
      "parametric_eq.band2_gain_db",
      "parametric_eq.band2_cutoff_freq",
      "parametric_eq.band2_q_factor",
      "parametric_eq.band3_gain_db",
      "parametric_eq.band3_cutoff_freq",
      "parametric_eq.band3_q_factor",
      "parametric_eq.high_shelf_gain_db",
      "parametric_eq.high_shelf_cutoff_freq",
      "parametric_eq.high_shelf_q_factor",
      "compressor.threshold_db",
      "compressor.ratio",
      "compressor.attack_ms",
      "compressor.release_ms",
      "compressor.knee_db",
      "compressor.makeup_gain_db",
      "distortion.drive_db"
    ]
  }
}
```

Generated by `dataset/build.py` in the [StemFlipper](https://huggingface.co/spaces/nakas/stemflipper) repo.