Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Hume AI — Speech Prosody Dataset

Interactive map: https://www.hume.ai/explore/speech-prosody-model

What this is

A dataset of ~8,000 short speech clips labelled with emotion scores, collected via Amazon Mechanical Turk and published by Alan Cowen, Dacher Keltner et al. (Hume AI / UC Berkeley).

Each clip is a neutral sentence spoken with a specific emotional prosody (tone, rhythm, timbre). MTurk raters listened and selected which emotions they heard; the scores are the proportion of raters who chose each emotion. Only emotions that passed a significance threshold are stored per clip — the rest are absent from that entry.

The dataset was originally published with 2,519 clips (Cowen et al., 2019). The version on the Hume AI website contains 7,988 clips.

How it was collected

  1. Recording: 100 actors from 5 cultures produced speech samples. They spoke lexically identical sentences (same words every time) while being induced into a target emotional state — so any emotional signal is purely prosodic.
  2. Rating: US and Indian participants on MTurk rated each clip on 48 emotion categories and several affective scales (valence, arousal, etc.).
  3. Analysis: Semantic space methods (PPCA across cultures) identified 12+ preserved dimensions of emotional prosody. Results visualized as a 2D t-SNE map.

Audio files

Google Drive: https://drive.google.com/drive/folders/1_KOPBVD6VsYa5pPDXcWIEgflg6fQAYVj

7,988 MP3 clips (~350 MB total). Filenames match the File field in the metadata (basename only).

Files

metadata.json          — full dataset, one object per clip (7,988 entries)
labels.csv             — same data in CSV format
scrape_and_download.py — script used to extract + download everything from hume.ai

Data format

Each entry has:

Field Type Description
File string Original S3 URL of the MP3
X int t-SNE x coordinate on the 2D emotion map
Y int t-SNE y coordinate on the 2D emotion map
Color hex string Dot color on the map (encodes dominant emotion cluster)
(emotion name) float 0–1 Proportion of raters who attributed that emotion

Only emotions above a significance threshold are included per clip. The full set of 46 possible emotion labels: Admiration, Adoration, Aesthetic_Appreciation, Amusement, Anger, Anxiety, Awe, Awkwardness, Boredom, Calmness, Concentration, Confusion, Contemplation, Contempt, Contentment, Craving, Desire, Determination, Disappointment, Disgust, Distress, Doubt, Ecstasy, Embarrassment, Empathic_Pain, Entrancement, Envy, Excitement, Fear, Guilt, Horror, Interest, Joy, Love, Nostalgia, Pain, Pride, Realization, Relief, Romance, Sadness, Satisfaction, Shame, Sympathy, Tiredness, Triumph.

Samples

Clip with high anger + blended aesthetics (dia973_utt12.mp3):

{
  "File": "https://mturkrecord.s3.amazonaws.com/targpros/dia973_utt12.mp3",
  "X": 270, "Y": 151, "Color": "#964E82",
  "Anger": 0.715, "Aesthetic_Appreciation": 0.538, "Calmness": 0.48,
  "Admiration": 0.326, "Adoration": 0.339, "Doubt": 0.353,
  "Distress": 0.247, "Joy": 0.251, "Disgust": 0.216
}

Clip dominated by anxiety (dia590_utt7.mp3):

{
  "File": "https://mturkrecord.s3.amazonaws.com/targpros/dia590_utt7.mp3",
  "X": 331, "Y": 512, "Color": "#906376",
  "Anxiety": 0.59, "Contemplation": 0.393, "Awkwardness": 0.362,
  "Determination": 0.225, "Interest": 0.207, "Adoration": 0.252
}

Papers

  • nihms-1518869.pdf — Cowen et al. (2019). The primacy of categories in the recognition of 12 emotions in speech prosody across two cultures. Nature Human Behaviour. The original study behind this dataset.
  • 1-s2.0-S136466132030276X-main.pdf — Cowen & Keltner (2021). Semantic Space Theory: A Computational Approach to Emotion. Trends in Cognitive Sciences. The theoretical framework.
  • 23Keltner-BasicEmotions.pdf — Keltner, Brooks & Cowen (2023). Semantic Space Theory: Data-Driven Insights Into Basic Emotions. Current Directions in Psychological Science.
  • Deep-learning-reveals-what-vocal-bursts-express-in-different-cultures.pdf — Brooks et al. (2022). Deep learning reveals what vocal bursts express in different cultures. Nature Human Behaviour. Related dataset (non-linguistic vocalizations, same lab).

About

Speech Prosody Dataset scraped from Hume AI website. Audios tagged with emotional classes such as amusement, anger, boredom, fear, ...

Resources

Stars

0 stars

Watchers

0 watching

Forks

Contributors

Languages