MLSpeech/hillenbrand_vowels

Hillenbrand Vowel Dataset

The Hillenbrand dataset contains recordings of American English vowels produced by speakers from four demographic groups: men, women, boys, and girls.
Each audio sample is accompanied by frame-level formant tracks (F1, F2, F3, F4) extracted every 10 ms.

This repository provides the dataset in a structure compatible with the Hugging Face datasets library for easy loading and processing.

---

Dataset Summary

• Sampling rate: 16 kHz
• Frame shift: 10 ms (formant extraction)
• Formants provided: F1, F2, F3, F4 (lists per frame)
• Groups: boys, girls, men, women
• Vowels: list of vowels present in the dataset
• Metadata file: metadata.csv describing path, group, vowel

---

Dataset Structure

Each dataset entry includes the following fields:

Field	Description
audio	The audio waveform (Hugging Face Audio type)
group	Speaker group (boys, girls, men, women)
vowel	The vowel label
formant_1	List of F1 values per 10 ms frame
formant_2	List of F2 values per frame
formant_3	List of F3 values per frame
formant_4	List of F4 values per frame

---

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("MLSpeech/hillenbrand")

print(ds[0])

# Example output:

{
  "file_name": {...},
  "group": "boys",
  "vowel": "ae",
  "formant_1": [...],
  "formant_2": [...],
  "formant_3": [...],
  "formant_4": [...],
}