BARI-EmoSpeech: A Dialect-Specific Speech Emotion Recognition Dataset for Barishal Bangla

Dataset Overview BARI-EmoSpeech is the inaugural public dataset for speech emotion recognition (SER) in the Barishal dialect of Bangla, addressing gaps in low-resource affective computing where standard models fail on dialectal traits like vowel elongations and intonation shifts. It includes 2,312 spontaneous audio utterances (1.67 hours total) across five emotions: Affection (500 samples), Happiness (407), Jealousy (605), Normal (500), and Sadness (300). Imbalance mirrors real-world elicitability, with Jealousy dominant and Sadness sparse. Collection Methodology Five native Barishal speakers (aged 22–45, gender-balanced) collected samples via culturally relevant prompts (e.g., festival recollections for Happiness). Recorded at 16 kHz on mobile devices in quiet settings (3–5 s average; median 2.28 s), data underwent dual annotation (Fleiss' Kappa κ = 0.82). Data Format and Access WAV files (mono, 16 kHz, 16-bit) in /train/{class}/ and /val/{class}/, plus CSV metadata (ID, duration, speaker, emotion). Includes F0/RMS violin plots.