A Tamil Emotional Speech Dataset
The first emotional speech dataset in Tamil — capturing the linguistic and emotional diversity of Sri Lankan Tamil for speech emotion recognition research.
Department of Computer Science & Engineering, University of Moratuwa · aaivu
EmoTa is distributed as a gated dataset on Hugging Face. Accept the usage agreement once and access is approved automatically — no waiting on email.
Create a free account or log in — your identity is tied to the agreement you accept.
Go to aaivu-labs/EmoTa and review the terms of use.
Complete the short access form stating your name, affiliation, and intended use.
Approval is automated. Once submitted, the audio and metadata are immediately available to download or stream.
The full corpus — 936 emotional utterances with speaker, region, and transcript metadata — behind a one-time usage agreement.
Open on Hugging FaceEmoTa comprises 936 recorded utterances from 22 native Sri Lankan Tamil speakers, each articulating 19 semantically neutral sentences across five primary emotions. Speakers represent dialects from the northern, eastern, western, and central provinces, and inter-annotator agreement reached a substantial Fleiss' κ of 0.74.
Balanced by gender — 11 male, 11 female — spanning four regional Sri Lankan Tamil dialects.
Anger, happiness, sadness, fear, and neutrality, elicited under a consistent protocol.
Semantically neutral prompts reduce lexical bias so the signal is emotion, not wording.
Captured in a controlled, soundproof environment with professional equipment.
Compact yet balanced coverage across every speaker, emotion, and sentence.
XGBoost and Random Forest baselines reach 0.91 and 0.90 F1 on emotion classification.
The emota_loader package on PyPI parses the corpus into typed samples with speaker, region, transcript, and emotion fields.
Install
$ pip install emota_loaderGet the audio first. Download the dataset from Hugging Face, then point the loader at its root directory.
Example output
Loaded 936 samples
Audio Path : EmoTa/19_18_ang.wav
Speaker ID : 19
Speaker Gender : male
Speaker Age : 25
Speaker Region : northern
Sentence ID : 18
Transcript : நான் உன்னை சந்திக்க வேண்டும்.
Emotion : angryUsage
from emota_loader import EmoTaDataset
# Point to the extracted root folder
dataset = EmoTaDataset(root_dir="path/to/EmoTa").samples
print(f"Loaded {len(dataset)} samples")
sample = dataset[0]
print(f" Audio Path : {sample.audio_path}")
print(f" Speaker ID : {sample.speaker_id}")
print(f" Speaker Gender : {sample.speaker_gender}")
print(f" Speaker Region : {sample.speaker_region}")
print(f" Sentence ID : {sample.sentence_id}")
print(f" Transcript : {sample.transcript}")
print(f" Emotion : {sample.emotion}")File structure
EmoTa/
├── happy/
├── sad/
├── angry/
├── fear/
└── neutral/
└── <spkID>_<senID>_<emo[:3]>.wavIf EmoTa supports your research, please cite the CHiPSAL 2025 paper.
@inproceedings{thevakumar-etal-2025-emota,
title = "{E}mo{T}a: A {T}amil Emotional Speech Dataset",
author = "Thevakumar, Jubeerathan and Thavarasa, Luxshan and
Sivatheepan, Thanikan and Kugarajah, Sajeev and
Thayasivam, Uthayasanker",
booktitle = "Proceedings of the First Workshop on Challenges in
Processing South Asian Languages (CHiPSAL 2025)",
month = jan,
year = "2025",
address = "Abu Dhabi, UAE",
publisher = "International Committee on Computational Linguistics",
url = "https://aclanthology.org/2025.chipsal-1.19/",
pages = "193--201"
}Built at the Artificial Intelligence and Autonomous Systems research group (aaivu), University of Moratuwa.
Acknowledgment. Thanks to all the volunteers who lent their voices to make EmoTa possible.
Special thanks to Braveenan Sritharan for his invaluable help and to Sajeev Kugarajah for his contribution to dataset collection.