csc-unipd/sonic-seasoning

Sonic Seasoning

Sonic Seasoning is a multi-source corpus of short music/sound clips, each perceptually rated for how strongly it evokes the five basic tastes — sweet, bitter, salty, sour, spicy — with additional temperature and emotion annotations on a subset. Every rating is normalized to [0, 1], and every row points to a single uniformly-encoded .wav file. It is the benchmark corpus for taste-from-audio prediction as a content-based music-information-retrieval task.

• 📄 Paper: Taste-aware music retrieval from audio embeddings — Matteo Spanio, Antonio Rodà (University of Padua), CBMI 2026.
• 💻 Code / models: github.com/CSCPadova/wav2taste · trained heads csc-unipd/wav2taste.
• 🧾 Ratings-only + retrieval queries (Zenodo companion): released under CC BY-NC 4.0.

At a glance

source	rows	rating columns provided
perceptual-validation	20	sweet, bitter, salty, sour, spicy
tasty-musicgen	100	sweet, bitter, salty, sour, hot, cold, happy, sad, anger, disgust, fear, surprise
annotated-corpus	257	sweet, bitter, salty, sour, spicy (sparse — see Coverage)
total	377	all of the above (unfilled cells are NaN)

Audio: mono, 44.1 kHz, 16-bit PCM .wav, one file per row at audio/<id>.wav.

Schema

column	type	filled for	description
id	string	all	<source>_<original_id>, sanitized to be filesystem-safe
audio_path	string	all	relative path to the audio file (audio/<id>.wav)
source	string	all	perceptual-validation / tasty-musicgen / annotated-corpus
paper	string	annotated-corpus only	original paper attribution (e.g. Mesz2011, Wang2017 (Spicy))
prompt	string	perceptual-validation + tasty-musicgen	for tasty-musicgen, the intended-taste category (bitter / salty / sour / sweet); for perceptual-validation, the curated food category
musical_genre	string	perceptual-validation only	original genre tag
split	string	all	train / val / test — see Splits
sweet	float	most rows	normalized to [0, 1]
bitter	float	most rows	normalized to [0, 1]
salty	float	most rows	normalized to [0, 1]
sour	float	most rows	normalized to [0, 1]
spicy	float	perceptual-validation + annotated-corpus	normalized to [0, 1] (tasty-musicgen rows are NaN)
hot, cold	float	tasty-musicgen only	temperature associations, [0, 1]
happy … surprise	float	tasty-musicgen only	six emotion associations, [0, 1]

Empty cells surface as NaN / None when loaded. Treat missing ratings as missing, not zero.

Sources & normalization

perceptual-validation — 20 stimuli

Curated real-music pieces evaluated by 49 participants on five 1–7 Likert scales (sweet, bitter, salty, sour, spicy). Per-stimulus mean scores were rescaled to [0, 1] via (x − 1) / 6. High inter-rater agreement makes this the held-out real-music generalization probe (assigned entirely to test) and the reference for the human-ceiling analysis in the paper.

tasty-musicgen — 100 stimuli

Text-to-music generations (25 per intended taste: bitter, salty, sour, sweet), rated by crowd participants along taste, temperature, and emotion categories. Values were normalized to [0, 1] upstream and copied as-is; the generation prompt is preserved in prompt. This source has no spicy rating (those cells are NaN).

annotated-corpus — 257 stimuli (subset)

Clips aggregated from 13 published sound–taste crossmodal studies. Ratings are frequency-based consensus values already in [0, 1]; the originating publication is preserved in paper for provenance and citation.

Coverage

Not every study measured every taste, so this source is sparse:

column	non-null rows
sweet	176
bitter	148
salty	177
sour	147
spicy	58

The 13 retained papers: Guedes2023 (Taste&Affect), Guetta2017, Knoeferle2015, Wang2015 (Whats), Wang2021 (Metacognition), Rodriguez2024, Reinoso2017, Mesz2011, Reinoso2020 (Blending), Reinoso2015JOSS, Wang2016 (Striking), Wang2017 (Spicy), Swahn2023. Other papers from the original meta-collection were excluded to keep a single consistent normalization.

Splits

A deterministic train / val / test partition (seed 42) in the split column:

split	rows	how it was built
train	269	80% of (annotated-corpus + non-test tasty-musicgen), stratified
val	68	20% of (annotated-corpus + non-test tasty-musicgen), stratified
test	40	all 20 perceptual-validation + 5 per intended-taste prompt of tasty-musicgen (4 × 5 = 20)

Rationale. perceptual-validation is held out entirely as a real-music generalization probe. tasty-musicgen contributes 20 test rows (5 per prompt) so the seven generated-only columns still have a held-out set. For train/val, stratification keys are tasty_<prompt> and annotated-corpus_spicy vs annotated-corpus_nospicy, keeping the sparse spicy ratings balanced across train and val. Clips from the same paper may appear in both train and val — if you need paper-level generalization, group by paper yourself.

Per-column non-null counts by split

	sweet	bitter	salty	sour	spicy	hot	cold	happy	sad	anger	disgust	fear	surprise
train	204	182	207	181	46	64	64	64	64	64	64	64	64
val	52	46	50	46	12	16	16	16	16	16	16	16	16
test	40	40	40	40	20	20	20	20	20	20	20	20	20

Loading

from datasets import load_dataset

ds = load_dataset("csc-unipd/sonic-seasoning", split="train")  # one HF split; use the `split` column
train = ds.filter(lambda r: r["split"] == "train")
val   = ds.filter(lambda r: r["split"] == "val")
test  = ds.filter(lambda r: r["split"] == "test")

Audio lives beside the table as .wav files referenced by audio_path. Resolve a clip:

from huggingface_hub import hf_hub_download
import soundfile as sf

row = test[0]
wav_path = hf_hub_download("csc-unipd/sonic-seasoning", row["audio_path"], repo_type="dataset")
audio, sr = sf.read(wav_path)   # mono, 44.1 kHz

Filter by source or to a single non-sparse task:

gen   = ds.filter(lambda r: r["source"] == "tasty-musicgen")
sweet = ds.filter(lambda r: r["sweet"] is not None)   # rows actually rated for sweet

Intended use

The corpus supports multi-task taste regression (one model, five outputs, masked loss over missing cells) and content-based retrieval (rank a corpus by predicted taste profile against food-descriptor queries). The paper evaluates ten frozen audio encoders under a shared head and operationalizes the predicted taste space as a retrieval index over a 309-item pool (train + test). Reference implementation and pretrained heads: github.com/CSCPadova/wav2taste.

Notes for downstream use

• Cross-source comparability is approximate. The three sources used different elicitation procedures (Likert means, crowd ratings, forced-choice frequencies). All values are rescaled to [0, 1], but a 0.5 from perceptual-validation and a 0.5 from annotated-corpus need not mean the same thing — use the source column to filter or weight.
• Per-column test sizes are uneven (spicy has 20 test rows vs sweet's 40). Report bootstrap confidence intervals when comparing models on the rarer columns.
• Audio is mono / 44.1 kHz / 16-bit, re-encoded from mixed original formats for uniformity. Only stimulus-level aggregates are included — no participant-level data.

Licensing

Released under CC BY-NC 4.0 (non-commercial, with attribution).

• Ratings, split, and compilation are authored by the dataset creators and are covered directly by CC BY-NC 4.0 (this mirrors the Zenodo ratings release).
• Audio provenance varies by source: perceptual-validation uses pre-existing music, tasty-musicgen is model-generated, and annotated-corpus aggregates stimuli from prior publications. The clips are redistributed here for non-commercial research use only; when using the annotated-corpus subset, also cite the originating studies recorded in the paper column.

Citation

@inproceedings{spanio2026taste,
  title     = {Taste-aware music retrieval from audio embeddings},
  author    = {Spanio, Matteo and Rod{\`a}, Antonio},
  booktitle = {Proceedings of the International Conference on Content-Based Multimedia Indexing (CBMI)},
  year      = {2026},
}

@misc{sonic_seasoning_dataset_2026,
  title        = {Sonic Seasoning: a multi-source perceptual dataset of taste-evoking sounds},
  author       = {Spanio, Matteo and Rod{\`a}, Antonio},
  year         = {2026},
  howpublished = {Hugging Face Hub},
  url          = {https://huggingface.co/datasets/csc-unipd/sonic-seasoning},
}

When using the annotated-corpus subset, please also cite the underlying source studies (attributions are preserved in the paper column).