dalmeow/crema-d_with_f0

CREMA-D with F0 Features

A tabular dataset derived from CREMA-D (Crowd-sourced Emotional Multimodal Actors Dataset) that augments the original corpus metadata with per-utterance fundamental frequency (F0 / pitch) statistics extracted using the CREPE deep learning pitch tracker.

7,442 utterances · 91 actors · 6 emotions · 12 sentences

Dataset Description

CREMA-D is a well-established acted emotional speech corpus in which 91 actors (ages 20–74, diverse in sex and race) each produced 12 sentences under 6 emotional conditions. This dataset does not redistribute the audio — it provides a ready-to-use CSV combining all original metadata with acoustic F0 features, suitable for statistical modelling or machine learning without needing to process the raw audio.

Source corpus

Cao, H., Cooper, D. G., Keutmann, M. K., Gur, R. C., Nenkova, A., & Verma, R. (2014). CREMA-D: Crowd-sourced Emotional Multimodal Actors Dataset. IEEE Transactions on Affective Computing, 5(4), 377–390. https://doi.org/10.1109/TAFFC.2014.2336244

F0 Extraction Methodology

F0 was extracted from each utterance using CREPE (medium model capacity) with Viterbi smoothing at a 16 kHz sample rate. Frames with a confidence score below 0.5 or with frequency ≤ 0 Hz were excluded as unvoiced. The following summary statistics were computed over the remaining voiced frames:

Feature	Description
meanF0	Mean F0 of voiced frames (Hz)
sdF0	Standard deviation of voiced F0 (Hz)
iqrF0	Interquartile range of voiced F0 (Hz) — robust spread measure
confidence	Mean CREPE confidence score for voiced frames

Dataset Structure

Columns

Column	Type	Description
Filename	string	Original WAV filename (e.g. 1001_IEO_NEU_XX.wav)
AudioPath	string	File path relative to the official CREMA-D GitLab mirror root (e.g. crema-d-mirror/AudioWAV/1001_IEO_NEU_XX.wav)
ActorID	string	4-digit actor identifier (1001–1091)
Age	int	Actor age in years (20–74)
Sex	string	Male or Female
Race	string	Caucasian, African American, Asian, or Hispanic
SentenceCode	string	3-letter sentence code (e.g. IEO)
Sentence	string	Full sentence text
EmotionCode	string	3-letter emotion code (e.g. ANG)
Emotion	string	Anger, Disgust, Fear, Happy, Neutral, or Sad
LevelCode	string	Intensity code: LO, MD, HI, or XX
Level	string	Low, Medium, High, or Unspecified
meanF0	float	Mean voiced F0 in Hz
sdF0	float	Standard deviation of voiced F0 in Hz
iqrF0	float	Interquartile range of voiced F0 in Hz
confidence	float	Mean CREPE confidence score (0–1)

Sentences

Code	Sentence
IEO	It's eleven o'clock
TIE	That is exactly what happened
IOM	I'm on my way to the meeting
IWW	I wonder what this is about
TAI	The airplane is almost full
MTI	Maybe tomorrow it will be cold
IWL	I would like a new alarm clock
ITH	I think I have a doctor's appointment
DFA	Don't forget a jacket
ITS	I think I've seen this before
TSI	The surface is slick
WSI	We'll stop in a couple of minutes

Usage

Python

from datasets import load_dataset

ds = load_dataset("dalmeow/crema-d_with_f0")
df = ds["train"].to_pandas()

# Mean F0 by emotion
df.groupby("Emotion")["meanF0"].mean().sort_values()

R

library(tidyverse)

df <- read_csv("https://huggingface.co/datasets/dalmeow/crema-d_with_f0/resolve/main/crema_d_with_f0.csv")

df |>
  ggplot(aes(log(meanF0), fill = Emotion)) +
  geom_density(alpha = 0.7) +
  scale_fill_brewer(palette = "Dark2") +
  labs(x = "log F0 (Hz)", title = "F0 distribution by emotion")

Bayesian regression example (R/brms)

library(brms)

f0_model <- brm(
  meanF0 ~ Emotion + (1 | ActorID),
  family = lognormal(),
  data = df,
  cores = 4
)

Limitations

• F0 values are summary statistics per utterance — temporal dynamics are not preserved.
• Acted speech may not generalise to spontaneous emotional expression.
• AudioPath values are relative to the CREMA-D GitLab mirror directory structure; adjust the root prefix to match your local clone.
• Utterances where no voiced frames exceeded the confidence threshold have NaN F0 values.

License

The underlying CREMA-D data is made available under the Open Database License (ODbL) 1.0; individual content within the database is covered by the Database Contents License (DbCL) 1.0. This derived dataset inherits those terms. The original audio files are not redistributed here.

Attribution

The F0 features (meanF0, sdF0, iqrF0, confidence) were extracted and added to this dataset by dalmeow using the CREPE pitch tracker. The underlying corpus metadata and audio recordings are the work of Cao et al. (2014).

Citation

Please cite the original CREMA-D paper when using this dataset:

@article{cao2014crema,
  title     = {{CREMA-D}: Crowd-sourced Emotional Multimodal Actors Dataset},
  author    = {Cao, Houwei and Cooper, David G. and Keutmann, Michael K. and Gur, Ruben C. and Nenkova, Ani and Verma, Ragini},
  journal   = {IEEE Transactions on Affective Computing},
  volume    = {5},
  number    = {4},
  pages     = {377--390},
  year      = {2014},
  publisher = {IEEE},
  doi       = {10.1109/TAFFC.2014.2336244}
}