BanglaSER: A speech emotion recognition dataset for the Bangla language

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Speech Processing

Background:

• Speech emotion recognition is crucial for human-computer interaction but challenging for low-resource languages like Bangla due to data scarcity.
• Existing research in Bangla speech emotion recognition is limited by the absence of comprehensive datasets.
• Bangla, despite being widely spoken, remains underrepresented in speech emotion recognition research.

Purpose of the Study:

• To introduce BanglaSER, a novel dataset for Bangla speech emotion recognition.
• To address the data scarcity issue for Bangla in the field of speech emotion recognition.
• To provide a valuable resource for training and evaluating Bangla speech emotion recognition models.

Main Methods:

• The BanglaSER dataset comprises 1467 audio recordings of 34 diverse speakers (17 male, 17 female).
• Recordings capture five basic emotions: angry, happy, neutral, sad, and surprise, with multiple repetitions.
• Speech data was collected using smartphones and laptops, ensuring a balanced and representative sample.

Main Results:

• The BanglaSER dataset offers a balanced distribution of emotional states and speaker demographics.
• It contains 1467 recordings suitable for training deep learning models like CNNs, LSTMs, and Transformers.
• The dataset is publicly available for research purposes.

Conclusions:

• BanglaSER is a significant contribution to Bangla speech emotion recognition, mitigating the low-resource language challenge.
• The dataset facilitates the development of more generalized and accurate Bangla speech emotion recognition systems.
• It is expected to spur further research and advancements in artificial intelligence for Bangla language processing.