Datasets:

NCSOFT
/

Designed-Vocalizations-Dataset

license: cc-by-4.0
pretty_name: Designed Vocalizations Dataset
language:
  - en
task_categories:
  - audio-to-audio
tags:
  - voice-conversion
  - sound-design
  - vocalization
  - speech
size_categories:
  - 100K<n<1M
configs:
  - config_name: raw
    data_files:
      - split: train
        path: data/raw/train-*.parquet
      - split: test
        path: data/raw/test-*.parquet
  - config_name: designed
    data_files:
      - split: train
        path: data/designed/train-*.parquet
      - split: test
        path: data/designed/test-*.parquet

Designed Vocalizations Dataset

Demo & audio samples

The Designed Vocalizations Dataset supports voice conversion for designed vocalizations — monster growls, robotic voices, and other sound-designed timbres — an area left underexplored by benchmarks that focus on natural human speech. It curates diverse raw vocal sources (speech and animal / non-linguistic sounds) and applies professional vocal-effects processing to produce corresponding effect-modified variants. A standardized test set with explicit seen/unseen splits over source-timbre groups and preset styles enables generalization evaluation under controlled conditions.

Dataset structure

Two configs — raw (source recordings) and designed (effect-processed clips) — each with a train and test split. One row per clip; the audio column holds the clip (44.1 kHz · 16-bit mono WAV, embedded).

raw — source recordings
- train — 5,654 training sources (linguistic 3,270 · non-linguistic 2,384)
- test — 120 evaluation sources (seen-source 60 · unseen-source 60)
designed — effect-processed clips
- train — 226,160 clips (40 training presets)
- test — 5,640 references, one per test source × preset (47 presets: seen 40 · unseen 7)

Train is non-parallel. seen-source = timbre groups present in training; seen-preset = presets also used in training. Overall: 237,574 clips.

Repo layout:

data/
    raw/        train-*.parquet, test-*.parquet
    designed/   train-*.parquet, test-*.parquet
metadata/
    preset_info.csv       # per preset: type, module chain
    preset_chains.json    # full DSP chain + parameters for in-house presets
    test_pairs.csv        # evaluation pairs + seen/unseen condition
assets/
README.md   LICENSE.md   NOTICE

Usage

from datasets import load_dataset

REPO = "NCSOFT/Designed-Vocalizations-Dataset"

# config = raw | designed, split = train | test
ds = load_dataset(REPO, "designed", split="test")   # 5,640 references

example = ds[0]
audio = example["audio"]            # {"array": np.ndarray, "sampling_rate": 44100, "path": str}
wav, sr = audio["array"], audio["sampling_rate"]
print(example["file_path"], example["preset"], example["source"])

Each example is one clip: audio decodes to a mono waveform (array) at 44.1 kHz, plus that config's metadata columns (see Columns). The four combinations:

load_dataset(REPO, "raw",      split="train")   # 5,654 training sources
load_dataset(REPO, "raw",      split="test")    # 120 evaluation inputs
load_dataset(REPO, "designed", split="train")   # 226,160 processed clips
load_dataset(REPO, "designed", split="test")    # 5,640 references

For a designed clip, source is the file_path of its originating/paired raw clip. See example.py for the full workflows — train (non-parallel: raw + designed pools) and test (parallel: source→reference pairs from metadata/test_pairs.csv).

Columns

Columns are the same across a config's train and test splits. Every row has audio (the clip) and file_path (original relative path, a stable identifier), plus:

raw

Field	Description
`origin`	`vctk`, `hifitts`, or `freesound`
`source_type`	Source category (e.g. `speech`, `animal`, `interjection`, `monster_mimic`)
`duration`	Clip length in seconds
`license`	`CC0 1.0`, `CC BY 3.0`, or `CC BY 4.0` (URLs in `NOTICE`)
`attribution_required`	`yes` if crediting the creator is required; `no` for CC0
`freesound_title` / `freesound_uploader` / `freesound_url` / `freesound_tags`	Freesound provenance (empty for speech)
`seen_in_train`	`yes` if the timbre was seen in training (all `yes` in `train`; mixed in `test`)

designed

Field	Description
`preset`	Preset applied (details in `preset_info.csv`)
`source`	The originating/paired source's `file_path` (in the `raw` config)

Evaluation conditions are in metadata/test_pairs.csv: seen_split is seen_to_seen, seen_to_unseen, unseen_to_seen, or unseen_to_unseen — first term = whether the source was seen in training, second = whether the preset was.

Preset metadata

Two lookup files under metadata/, keyed by preset.

preset_info.csv

Preset characteristics, one row per preset.

Field	Description
`preset`	Preset identifier (matches the `preset` column in the `designed` config)
`preset_type`	`in-house` (self-built) or `dehumaniser_builtin` (built on a Dehumaniser stock preset)
`seen_in_train`	`yes` if used in the train split, `no` if test-only (the 7 unseen presets: `preset_26`–`preset_30`, `preset_46`–`preset_47`)
`num_dsp_modules`	Total number of DSP module instances in the chain (an `xN` module counts as N)
`dsp_modules`	Readable module chain; `xN` denotes N instances of a module

preset_chains.json

Full, reproducible DSP chain and parameters for the in-house presets (17 of them). Built-in Dehumaniser presets are not detailed here (only their effect list, in preset_info.dsp_modules). Nested JSON keyed by preset name:

{
  "<preset>": {
    "summary": "[ Pitch Shifting 1 ∥ Ring Modulator ∥ (Granular → Pitch Shifting 2) ] → Reverb → EQ",
    "chain": [
      { "plugin": "Dehumaniser", "branches": [
          [ { "effect": "Pitch Shifting 1", "params": {...} } ],
          [ { "effect": "Ring Modulator", "params": {...} } ],
          [ { "effect": "Granular", "params": {...} },
            { "effect": "Pitch Shifting 2", "params": {...} } ]
      ] },
      { "effect": "Reverb", "params": {...} },
      { "effect": "EQ", "params": { "bands": [ {"freq": "1598Hz", "gain": "+8.3dB"} ] } }
    ]
  }
}

chain runs in series in list order. Each stage, and the summary field:

plugin — a Dehumaniser stage: a container whose branches run in parallel, with the effects inside a branch running in series.
effect — a single-effect stage: mostly Dehumaniser's internal modules (Pitch Shifting, Granular, Ring Modulator, …); in-house presets may also add Cubase stock effects (Compressor, EQ, Reverb, Limiter, Distortion).
params — an effect's settings, usually name/value pairs; for EQ, a bands list of freq/gain pairs.
summary — the chain on one line: → series, ∥ parallel, [ … ] a summed parallel block, ( … ) a serial sub-chain.

Visualization of the example preset shown above.

Licensing

The dataset is made available under CC BY 4.0 (see LICENSE.md) for the dataset compilation and the authors' original contributions, including metadata and original annotations. Designed audio clips may incorporate or be adaptations of third-party source sounds and therefore remain subject to the applicable source licenses identified per clip.

Those sources keep their original licenses — VCTK (CC BY 4.0), HiFi-TTS (CC BY 4.0), and Freesound (CC0 / CC BY 3.0 / CC BY 4.0). Their attribution/notice requirements are in NOTICE, with per-clip license, Freesound title, uploader, and source link carried as columns in the raw config. These must be kept when redistributing.

Citation

If you use this dataset, please cite:

S. Lee, M. Kang, Y. Lee, W. Min, C. Lee, and N. Cho, "Designed Vocalizations Dataset: Sound-Designed Human and Animal Voices for Non-human Voice Conversion," in Proc. Interspeech 2026.

@inproceedings{lee2026designed,
  title     = {Designed Vocalizations Dataset: Sound-Designed Human and Animal Voices
               for Non-human Voice Conversion},
  author    = {Lee, Seolhee and Kang, Minsu and Lee, Yangsun and Min, Woosun and
               Lee, Choonghyeon and Cho, Namhyun},
  booktitle = {Proc. Interspeech 2026},
  year      = {2026}
}