Datasetpublic

OlhaHavryliuk/UA-SER

UA-SER: Ukrainian Speech Emotion Recognition Corpus A labelled Ukrainian emotional speech corpus of 952 clips across four emotion classes, collected and annotated for the purpose of training and evaluating Speech Emotion Recognition (SER) models on Ukrainian. Dataset Summary Ukrainian is a low-resource language with no publicly available emotional speech dataset. UA-SER fills this gap by providing short naturalistic utterances labelled by three native Ukrainian… See the full description on the dataset page: https://huggingface.co/datasets/OlhaHavryliuk/UA-SER.

sourceHugging Facecc-by-4.0updated 5mo agoView on Hugging Face

0likes20downloads

Dataset Card

UA-SER: Ukrainian Speech Emotion Recognition Corpus

A labelled Ukrainian emotional speech corpus of 952 clips across four emotion classes, collected and annotated for the purpose of training and evaluating Speech Emotion Recognition (SER) models on Ukrainian.

Dataset Summary

Ukrainian is a low-resource language with no publicly available emotional speech dataset. UA-SER fills this gap by providing short naturalistic utterances labelled by three native Ukrainian annotators via majority vote.

Property	Value
Language	Ukrainian (`uk`)
Clips	952
Emotion classes	4 (angry, happy, neutral, sad)
Speakers	400
Total duration	~33 min
Mean clip duration	2.07 s
Annotation	3 native annotators, majority vote
Inter-annotator agreement	Fleiss κ = 0.46
Split strategy	Speaker-disjoint 80/20

Emotion Distribution

Emotion	Total	Train	Test
Angry	259	210	49
Happy	244	200	44
Sad	227	181	46
Neutral	222	180	42
Total	952	771	181

Gender Distribution

Gender	Count
Female	593
Male	359

Gender is encoded in the filename suffix: _0.wav = male, _1.wav = female.

Dataset Structure

clips/          # 952 WAV files, 16 kHz mono
dataset.csv     # metadata (see below)

`dataset.csv` columns

Column	Description
`filename`	WAV filename (e.g. `s_001_1.wav`)
`emotion`	Label: `angry`, `happy`, `neutral`, `sad`
`duration_seconds`	Clip duration in seconds
`text`	Transcript of the utterance (Ukrainian)
`speaker_id`	Anonymised numeric speaker identifier
`split`	`train` or `test`

Filename convention

{source_prefix}_{clip_id}_{gender}.wav

—source_prefix — corpus source identifier
—clip_id — zero-padded sequential index
—gender — 0 (male) or 1 (female)

Train / Test Split

The split is speaker-disjoint: no speaker appears in both train and test. Speakers were assigned greedily to maximise balance of emotion and gender distributions between the two sets (target 80/20 ratio).

Annotation

Each clip was independently labelled by three native Ukrainian speakers. The final label is the majority vote (at least 2 of 3 agree). Clips where no majority was reached were excluded.

Inter-annotator agreement: Fleiss κ = 0.46 (moderate agreement), consistent with other naturalistic emotion corpora.

Audio Format

—Format: WAV, 16-bit PCM
—Sample rate: 16 000 Hz
—Channels: mono
—Duration range: 0.52 – 5.50 s

Usage

python

import pandas as pd
import librosa

df = pd.read_csv("dataset.csv")
test = df[df["split"] == "test"]

y, sr = librosa.load(f"clips/{test.iloc[0]['filename']}", sr=16000)