README.md

---
pretty_name: WHiSER
tags:
- audio
- speech
- emotion
task_categories:
- audio-classification
- other
task_ids:
- audio-emotion-recognition
language:
- en
license: other
size_categories:
- 1K<n<10K
dataset_summary: >
  WHiSER is a multi-annotator speech emotion corpus providing categorical labels and
  dimensional V-A-D (arousal, valence, dominance; SAM 1–7) ratings for short speech clips.
  This repo contains raw WAV files and rich annotations in raw (labels.txt), consensus, and detailed formats (CSV/JSON).
papers:
- https://www.isca-archive.org/interspeech_2024/naini24_interspeech.pdf

extra_gated_prompt: "You agree to not use the dataset to conduct experiments that cause harm to human subjects."
extra_gated_fields:
  Company: text
  Country: country
  Specific date: date_picker
  I want to use this dataset for:
    type: select
    options: 
      - Research
      - Education
      - label: Other
        value: other
  I agree to use this dataset for non-commercial use ONLY: checkbox
---

# Dataset Description

## Motivation
- Enable both categorical emotion recognition and dimensional affect regression on speech clips with multiple crowd-sourced annotations per instance.
- Study annotator variability, label aggregation, and uncertainty.

## Composition
- Audio: WAV files in `wavs/`
- Annotations:
  - Raw: `WHiSER/Labels/labels.txt` (per-file summary + multiple worker lines)
  - Aggregated: `WHiSER/Labels/labels_consensus.{csv,json}`
  - Per-annotator: `WHiSER/Labels/labels_detailed.{csv,json}`

# Supported Tasks
- Categorical speech emotion recognition (primary label)
- Dimensional affect regression (arousal/valence/dominance; 1–7 SAM scale)
- Research on annotation aggregation and inter-annotator disagreement

# Dataset Structure

## Files
- `wavs/`: raw audio clips (e.g., `006-040.1-2_13.wav`)
- `WHiSER/Labels/labels.txt`
  - Header (one per file): `<filename>; <agg_code>; A:<float>; V:<float>; D:<float>;`
  - Worker lines: `WORKER<ID>; <primary>; <secondary_csv>; A:<float>; V:<float>; D:<float>;`
- `WHiSER/Labels/labels_consensus.{csv,json}`
  - Aggregated primary emotion (and secondary/attributes when applicable), gender, speaker, split (if defined)
- `WHiSER/Labels/labels_detailed.{csv,json}`
  - Individual annotations per worker with primary, secondary (multi-label), and V-A-D

## Recommended Features 
- `id`: string (e.g., `006-040.1-2_13`)
- `file`: string (e.g., `wavs/006-040.1-2_13.wav`)
- `audio`: Audio feature
- `agg_label_code`: string in `{A,S,H,U,F,D,C,N,O,X}`
- `agg_primary`: normalized primary label (e.g., `Happy`)
- `vad_mean`: `{ arousal: float32, valence: float32, dominance: float32 }` (1–7)
- `secondary`: sequence of strings (from the secondary label set)
- Optional metadata: `split`, `speaker`, `gender`
- `annotations` (from detailed): sequence of {
  `worker_id`: string,
  `primary`: string,
  `secondary`: sequence<string>,
  `vad`: { `arousal`: float32, `valence`: float32, `dominance`: float32 }
}

# Label Schema

## Primary consensus codes
- A: Angry
- S: Sad
- H: Happy
- U: Surprise
- F: Fear
- D: Disgust
- C: Contempt
- N: Neutral
- O: Other
- X: No agreement (no plurality winner)

## Secondary labels (multi-select)
- Angry, Sad, Happy, Amused, Neutral, Frustrated, Depressed, Surprise, Concerned,
  Disgust, Disappointed, Excited, Confused, Annoyed, Fear, Contempt, Other

## Dimensional affect
- Self-Assessment Manikin (SAM) 1–7
- Valence (1 very negative; 7 very positive)
- Arousal (1 very calm; 7 very active)
- Dominance (1 very weak; 7 very strong)

# Example Instance 
- `id`: `006-040.1-2_13`
- `file`: `wavs/006-040.1-2_13.wav`
- `agg_label_code`: `H`
- `vad_mean`: `{ arousal: 4.2, valence: 4.8, dominance: 5.0 }`
- `annotations` (subset):
  - `{ worker_id: WORKER00014325, primary: Sad, secondary: [Sad, Concerned], vad: {A:2, V:4, D:4} }`
  - `{ worker_id: WORKER00014332, primary: Happy, secondary: [Happy, Concerned], vad: {A:4, V:6, D:6} }`


# Usage

## Loading from Parquet

If you've converted the dataset to nested Parquet format using the provided script, you can load it with the Hugging Face `datasets` library:

```python
from datasets import Dataset
import pyarrow.parquet as pq

# Load the Parquet file
table = pq.read_table("parquet/whiser_nested.parquet")
ds = Dataset(table)

print(f"Dataset size: {len(ds)}")
print(f"Features: {ds.features}")

# Access a single example
example = ds[0]
print(f"ID: {example['id']}")
print(f"Consensus label: {example['agg_primary']}")
print(f"VAD mean: A={example['vad_mean']['arousal']:.2f}, V={example['vad_mean']['valence']:.2f}, D={example['vad_mean']['dominance']:.2f}")
print(f"Secondary labels: {example['secondary']}")
print(f"Number of annotators: {len(example['annotations'])}")

# Access audio bytes (embedded in Parquet)
audio_bytes = example['audio_bytes']
sample_rate = example['sample_rate']
print(f"Audio: {len(audio_bytes)} bytes, {sample_rate} Hz")

# Iterate through per-annotator annotations
for ann in example['annotations']:
    worker = ann['worker_id']
    primary = ann['primary']
    vad = ann['vad']
    print(f"  Worker {worker}: primary={primary}, A={vad['arousal']}, V={vad['valence']}, D={vad['dominance']}")
```

## Loading Audio

To decode audio from the embedded bytes:

```python
import io
import soundfile as sf

example = ds[0]
audio_bytes = example['audio_bytes']

# Decode audio
audio_data, sr = sf.read(io.BytesIO(audio_bytes))
print(f"Audio shape: {audio_data.shape}, sample rate: {sr}")
```

# Reference

```bibtex
@inproceedings{Naini_2024_2, 
	author={A. {Reddy Naini} and L. Goncalves and M.A. Kohler and D. Robinson and E. Richerson and C. Busso}, 
	title={{WHiSER}: {White House Tapes} Speech Emotion Recognition Corpus},
	booktitle={Interspeech 2024}, 
	volume={},
	year={2024}, 
	month={September}, 
	address =  {Kos Island, Greece},
	pages={1595-1599}, 
	doi={10.21437/Interspeech.2024-1227},
}
```